Automated Feature Engineering, AutoML, and Decision-Focused Learning for Improved Energy Consumption Forecasting

knowledgement and profound recognition that human knowledge,

• Human knowledge, despite its ambition, rigour, and achievement, is described as but a faint glimmer.Source: knowledgement and profound recognition… · p. 5

mer of light within a vast surrounding darkness of uncertainty

• However boldly humanity advances as a questioning species, vast immensities remain beyond the horizon of knowing and comprehension.Source: mer of light within a vast surrounding… · p. 5

5.4.1 Performance Overview of Predict-Then-Optimise Compared

• The section is titled as a performance overview comparing predict-then-optimise with decision-focused learning.The supplied passage contains only a contents fragment, not performance results.Source: §5.4.1 · p. 9

5.4.2 Impact of Automated Feature Engineering on Predict-Then-Optimise and Decision-Focused Learning Methods with Fea-

• A following subsection is titled as an analysis of automated feature engineering’s impact on predict-then-optimise and decision-focused learning methods.The supplied passage contains only a fragment of the subsection heading.Source: §5.4.2 · p. 10

5.4.3 Feature Importance Analysis Across Predict-Then-Optimise

• Another listed subsection concerns feature importance analysis across predict-then-optimise methods.No feature-importance findings are included in the supplied passages.Source: §5.4.3 · p. 10

3.3 3D schematic representation of the investigated FDCSs. Key components indicated include (i) evaporator unit, responsi-

• FDCS 1 has higher variability in electricity, temperature, and humidity than FDCS 2, suggesting greater sensitivity to internal or external factors.The comparison uses coefficient of variation to quantify relative variability.Source: §3.3 · p. 19
• Lag and nested-window features contribute most to feature importance across the energy datasets, consistent with capturing temporal dependencies and multi-scale behaviour.The passage relates these features to lagged effects such as thermal inertia and patterns captured through nested windows.Source: §3.3 · p. 22
• The proposed framework uses AutoEnergy features to jointly forecast next-day electricity prices and demand while optimising battery charging and discharging decisions.The BESS-specific formulation could be adapted to similar energy storage systems by revising constraints, variables, and objectives.Source: §3.3 · p. 22
• Test-regret results support DFL with automated feature engineering in this experimental setting, but do not establish that DFL is universally superior.The comparison covers PTO and DFL methods (SPO+ and DBB), with and without automated feature engineering.Source: §3.3 · p. 24

1.1 Background and Motivation

• Manual feature engineering is time-consuming, prone to human error, and dependent on domain expertise, while automated features target generalisable knowledge rather than replacing site-specific expertise.The thesis distinguishes broadly useful patterns such as temporal cycles from local factors such as occupancy, equipment controls, and maintenance.Source: §1.1 · pp. 31–33
• AutoML automates model selection and tuning but often leaves data preparation and feature generation to practitioners, while existing automated feature engineering methods were not designed for energy forecasting.Deep learning can learn representations, but its interpretability and data-size limitations can be challenging in real-world settings; TSfresh and FeatureTools may miss energy-specific temporal patterns and consumption characteristics.Source: §1.1 · p. 33
• Forecasting accuracy alone does not ensure better practical outcomes, so automated feature engineering should also be evaluated by downstream decision quality and operational cost or utility.The passages motivate evaluation in Predict-Then-Optimise frameworks and Decision-Focused Learning [49] [50].Source: §1.1 · p. 33, §1.2 · p. 34

1.2. RESEARCH AIM AND OBJECTIVES

• The thesis hypothesises that combining domain-specific automated feature engineering with AutoML and decision-focused learning can reduce domain-knowledge dependence, improve forecasts, and optimise downstream tasks.Source: §1.2 · p. 34

1.2 Research Aim and Objectives

• The thesis aims to build ECF models applicable across diverse energy systems while maximising forecasting accuracy and optimising downstream tasks, with reduced manual, dataset-specific feature engineering.The stated reduction in domain knowledge is scoped to feature engineering and its associated selection choices.Source: §1.2 · p. 34

gate the role of domain knowledge in FE for such forecasting

• RO1 establishes baseline ECF models and a detailed pipeline, including historical electricity consumption and weather data as a real-world case study.Source: §1.2 · p. 35, gate the role of domain knowledge in… · p. 35
• RO1 also investigates domain-specific features, feature-selection techniques, algorithms, hyperparameter tuning, feature importance, and dataset-size effects on model performance.The intended output is a comprehensive ECF pipeline that helps practitioners reduce the time needed to implement domain-knowledge-based features.Source: gate the role of domain knowledge in… · p. 35

streamlining ML model development by addressing the most

• RO2 develops automated feature engineering for ECF by addressing which features to generate and how to automate feature selection.The proposed algorithm is intended to produce human-readable, traceable features from explicit transformations, while reducing time and expertise barriers and improving AutoML performance.Source: gate the role of domain knowledge in… · p. 35, streamlining ML model development by… · p. 35, §1.3 · p. 36

3. RO3: To leverage the proposed AFE algorithm to opti-

• RO3 extends AFE evaluation to downstream optimisation, aims to improve DFL methods, and tests practical viability on a novel real-world dataset and settings.The objective is to make AFE decision-aware so it can deliver operational value in ECF applications.Source: §3 · p. 36

1.3 Contribution and Novelty

• The ECF pipeline supports one-week-ahead hourly forecasts and small training datasets, while investigating domain knowledge in FE, eight FS methods, feature importance, and dataset-size effects.Its forecast horizon comprises the next 168 hours.Source: §1.3 · pp. 36–37

ECF problems to minimise reliance on domain knowledge

• AutoEnergy generates interpretable features from timestamps and past consumption, integrates with AutoML, and is evaluated across eighteen datasets spanning diverse energy domains.The thesis reports that it consistently outperforms baseline AutoML and achieves greater average reductions in forecasting errors and processing time than existing FE methods.Source: §1.3 · p. 37, ECF problems to minimise reliance on… · p. 37

demand and price over a 24-hour horizon while leveraging

• The RO3 framework evaluates AutoEnergy in a real-world Battery Energy Storage System problem using PTO and DFL, assessing decision quality with regret rather than forecasting metrics.The chapter examines how AFE affects operational decision quality, particularly under data scarcity.Source: demand and price over a 24-hour… · p. 38, §1.4 · p. 39

2.1.2.1 Feature Generation and Extraction

• ECF features encode operating schedules with timestamp indicators, represent cyclical time smoothly with sine and cosine transforms, and capture historical dependence through lags and rolling summaries.These feature types provide signals for routine patterns, periodicity, autocorrelation, local level, and variability.Source: Calendar and Schedule Indicators · p. 49, Smooth Cyclic Time Encodings · p. 49, Historical Dependence and… · p. 50

2.1.2.2 Feature Selection

• Feature selection can remove noisy or redundant inputs but is context-dependent, while manual FE is time-consuming and expertise-dependent, motivating automated feature-generation and selection strategies.AFE searches combinations of feature-processing operations and uses optimisation to identify an effective strategy.Source: §2.1.2.2 · p. 51, §2.1.3 · p. 52

2.1.3 Automated Feature Engineering

• Existing feature-generation approaches pose trade-offs: autoencoders and statistical transforms can obscure feature meaning, while heuristic methods can be costly or overfit.The chapter benchmarks heuristic approaches because they are rule-based, traceable, applicable to energy forecasting without extra training or human intervention, and potentially easier to integrate with AutoML.Source: §2.1.3 · pp. 52–54, §2.2 · p. 55

2.2. MACHINE LEARNING ALGORITHMS FOR ENERGY CONSUMPTION FORECASTING

• Energy forecasting has used traditional, neural, hybrid, ensemble, and multi-task models, with algorithm choice depending on dataset challenges, computational efficiency, interpretability, and researcher preference.Examples include KNN, neural networks, CNN-LSTM hybrids, XGB, RF, and multi-task models.Source: ergy Consumption Forecasting · p. 55, §2.3 · p. 56

2.3. AUTOMATED MACHINE LEARNING

• General AutoML automates preprocessing, model selection, and tuning but often assumes features are ready, leaving interpretable, domain-relevant feature generation to practitioners.AutoML and automated feature engineering address complementary stages of the pipeline.Source: §2.3 · pp. 56–57

2.4.1 Concept and Rationale

• Sequential predict-then-optimise can propagate forecast errors and misalign training with downstream costs; decision-focused learning instead trains forecasts against optimisation outcomes using regret-based losses.In battery scheduling, even small forecast shifts can change charging and discharging decisions and increase decision regret.Source: §2.4.1 · pp. 58–59, §2.4.2 · p. 63

2.4.2 Key Decision-Focused Learning Methods

• Automated feature engineering remains untested for improving decision-focused learning in data-constrained real-world settings, where manual feature design is expert-dependent and time-consuming.The cited work identifies small-scale operational datasets as a validation gap for decision-focused learning.Source: §2.4.1 · p. 60, §2.4.2 · p. 64

2.4.1 Concept and Rationale

• The proposed framework jointly forecasts electricity demand and prices and optimises battery storage using regret-based objectives, integrating domain-specific automated feature engineering and validated on real UK property data.It is designed for small datasets and compares decision-focused learning with traditional predict-then-optimise approaches.Source: §2.4.1 · p. 60

STORAGE FACILITIES

• The chapter establishes a practical forecasting pipeline from two real-world electricity-consumption and weather datasets, examining domain features, feature selection, algorithms, tuning, feature importance, and dataset size.Its findings provide an empirical foundation for subsequent feature-engineering automation.Source: STORAGE FACILITIES · p. 65

3.1 Introduction

• For food and drink cold storage, the work proposes a small-data pipeline for one-week-ahead hourly forecasts of electricity use, indoor temperature, and humidity, validated on two Nottingham datasets.These variables reflect the sector’s stringent temperature and humidity requirements for food safety and quality.Source: §3.1 · pp. 67, 69
• The experiments also investigate how eight feature-selection methods and varying dataset volumes affect FDCS forecasting, including the estimation of a minimum dataset size.The introduction notes that feature-selection findings may be data-dependent and that limited collection time, cost, or operational constraints can restrict available data.Source: §3.1 · pp. 69–70

3.2 Methodology

• The forecasting pipeline predicts electricity consumption, indoor temperature, and humidity one week ahead and examines eight feature-selection methods across filter, embedded, wrapper, and hybrid approaches.Linear Regression guided RFE and Sequential Forward Selection to keep feature selection consistent across downstream regressors; MAE was the primary comparison metric.Source: §3.1 · p. 69, §3.2 · p. 70, §3.2.3 · pp. 81–82

3.2.1 Data Collection

• The study collected electricity, temperature, and humidity data from two Nottingham FDCSs and used historical weather, with chronological 70%/15%/15% training, validation, and test splits.The two sites provide a purposeful robustness assessment across different operating regimes, but do not span all ECF configurations; the chapter investigates domain knowledge in depth, while broader generality is addressed elsewhere in the thesis.Source: §3.2.1 · pp. 72–73

3.2.2 Feature Extraction

• The feature set uses working-hour features, cyclical time transforms, selected historical lags, and rolling-window statistical features computed from past observations to capture temporal patterns without look-ahead bias.For the 168-hour forecast horizon, the ten highest-correlation lags were selected from lags 168–336.Source: §3.2.2 · p. 77, §3.2.2.1 · p. 78, §3.2.2.2 · p. 78

3.3.1 Performance Evaluation

• 150.75 and 157.79 MAE were the lowest FDCS 1 electricity test errors for XGBR and RFR, respectively, even without feature selection.Across both FDCSs, XGBR and RFR consistently achieved top-tier performance; in FDCS 2, embedded-lasso selection accompanied the lowest indoor-temperature errors for LSTM, MTL, RFR, and XGBR.Source: Figure 3.7 · p. 87, Figure 3.8 · p. 88, §3.3.1 · pp. 90–91
• Hybrid filter-plus-embedded selection most improved LSTM, MLP, KNNR, and MTL in FDCS 1, while neural-network performance varied across selection methods and dataset sizes.The study used six algorithms, including KNN, RF, XGB, MLP, LSTM, and multi-task learning, with grid-searched hyperparameters evaluated on a held-out test set.Source: §3.2.4 · pp. 82, 84, §3.3.1 · pp. 90–91
• FDCS 1 variables were more predictable than FDCS 2 variables, but greater predictability indicates forecasting value rather than energy efficiency.The authors relate this difference to more consistent daily patterns in FDCS 1 and caution that a predictable system may still consume more energy.Source: Figure 3.5 · p. 75, §3.3.1 · p. 92

3.3.2 Evaluating Feature Importance Using SHAP

• SHAP identified hour of day as most critical for FDCS 1 electricity and temperature predictions, while FDCS 2 relied on lagged, cyclical, rolling-statistic, and weather features.Four of FDCS 1’s top ten electricity features were extracted features; outdoor temperature influenced both systems, and FDCS 2 electricity predictions showed greater feature-impact variability.Source: §3.3.2 · pp. 93–96, Figure 3.10 · p. 95

3.3.3 Dataset Size Implications

• Electricity MAE improvements were most notable at 1344 hours in FDCS 1 and 1680 hours in FDCS 2, after which performance tended to plateau.Temperature stabilized around 1560 hours; humidity MAE decreased with more data but fluctuated. The analysis used XGBR and RFR, beginning with one day of hourly data and expanding in daily increments, and was computationally expensive; findings are limited to the available datasets.Source: Figure 3.12 · p. 97, §3.3.3 · pp. 98–99

3.4 Summary

• RFR and XGBR had the lowest MAE in both FDCS datasets; hybrid feature-selection methods generally improved performance, while wrapper methods were computationally expensive.SHAP analysis indicated that extracted features noticeably affected model outputs.Source: §3.4 · p. 100

4.2.2 The Proposed Feature Engineering Method

• AutoEnergy applies explicit, traceable rules to timestamps and past consumption, generating temporal and cyclical encodings, lags, and rolling-window statistics with statistical candidate selection.Lag and rolling features use only historical values available before the prediction time.Source: CONSUMPTION FORECASTING WITH AUTOML · pp. 101–102, §4.2.2 · p. 108, §4.2 · p. 110

4.1 Introduction

• AutoML automates model selection and tuning, but typically leaves feature engineering and domain-knowledge integration to practitioners, motivating AutoEnergy’s automated, domain-aware features.The work frames reduced manual effort as automation and reproducibility rather than directly measured time saved.Source: §4.1 · pp. 103–105

CONSUMPTION FORECASTING WITH AUTOML

• Across eighteen diverse energy datasets, AutoEnergy achieved superior forecasting-error reductions and computational efficiency versus established feature-engineering benchmarks.The chapter also examined its integration with TabPFN, reporting significant forecasting-error reductions across test sets.Source: CONSUMPTION FORECASTING WITH AUTOML · p. 102, §4.1 · pp. 105–106

4.2.2.1 Temporal, Sine and Cosine Transform-Derived Features

• Sine and cosine encodings represent daily and weekly cycles while preserving the circular adjacency of time, so hours 23 and 0 remain neighbours.The encoding gives each hour unique coordinates and constant distances between consecutive hours, avoiding the discontinuity of ordinal hour encoding.Source: §4.2.2.1 · p. 111, §4.2 · p. 113

4.2.2.2 Lags and Nested Window Features

• AutoEnergy uses Kendall’s tau and a p-value threshold of 0.05 to select significant target lags, retaining at most the ten significant lags with lowest p-values.Kendall’s tau is rank-based, distribution-free, and described as more robust to outliers than Pearson correlation; the lag-selection approach is inspired by FRESH [176].Source: §4.2.2.2 · pp. 114–115
• For rolling statistics, AutoEnergy tests nested windows up to one-third of the series length, retains sizes with p < 0.05, and balances multi-scale capture against redundant or costly windows.The one-third cap was empirically assessed against one-quarter, one-half, and full-sequence thresholds.Source: §4.2.2.2 · pp. 115–116

4.2.3 Considerations

• Generated features use only data available at time ti to prevent look-ahead bias, while lags and rolling statistics are intended to make historical dependencies and usage patterns interpretable.Rolling means, standard deviations, extrema, skewness, and kurtosis can represent baseline use, volatility, extremes, and distribution shape.Source: §4.2.3 · pp. 118–119

4.3.1 Datasets

• The evaluation used eighteen real-world datasets spanning residential, industrial, cold-storage, regional-use, and renewable energy settings, comparing AutoEnergy with TS and FT automated feature-engineering benchmarks.The datasets include both univariate and multivariate series with differing sample sizes and temporal resolutions.Source: §4.3.1 · pp. 119–120, Table 4.2 · p. 120, §4.3.2 · p. 122

4.3.3 Experimental Procedure

• Datasets were split chronologically into 80% training and 20% testing; AutoGluon was evaluated without features and after each FE method, while TabPFN comparisons were limited to eight datasets meeting its constraints.The no-feature AutoGluon run provided a baseline, and TabPFN required datasets with at most 10,000 samples and 500 features.Source: §4.3.3 · pp. 125–126

4.3.4 Evaluation Metrics and Statistical Tests

• Methods were assessed using normalized RMSE and FE processing time, with Friedman aligned-ranks tests, Bonferroni post hoc comparisons, and Wilcoxon signed-rank tests at α = 0.05.The study used nRMSE for cross-dataset error comparisons and processing time to evaluate computational efficiency.Source: §4.3.4 · pp. 126–127

4.4.1 Impact of AutoEnergy on AutoML Predictivity

• 83.22% average nRMSE reduction, ranging from 28.22% on Appliances to 98.39% on PJME, versus AutoGluon without feature engineering.Prediction stability improved by 53.69%, with standard deviation falling from 0.0773 to 0.0358.Source: §4.4.1 · p. 128, Table 4.3 · p. 130
• On six of eight TabPFN-compatible datasets, AutoEnergy produced the lowest forecasting errors among the compared feature-engineering methods; TSMin and FT won on Steel and Victoria, respectively.AutoEnergy supplied a TabPFN-compatible tabular feature matrix as preprocessing while TabPFN remained unchanged as the predictor.Source: §4.4.1 · pp. 128–129, Table 4.4 · p. 132

4.4.2 Comparison with Benchmark Methods

• AutoEnergy’s advantage was less pronounced for skewed Appliances and Steel datasets with exogenous features, though it had the lowest nRMSE on Steel and remained competitive on Appliances.Existing informative variables may reduce the marginal benefit of temporal features; FT performed best on Appliances.Source: §4.4.2 · pp. 133–134
• The one-third nested-window cap had the lowest average nRMSE (0.1758) and second-fastest processing time, with minimum error on 6 of 18 datasets.The half cap achieved minimum error on 7 datasets but was considerably slower; results were data-dependent, while the one-third cap remained near-optimal across diverse datasets and reduced the need for expert input.Source: §4.4.2 · pp. 134–135, Table 4.5 · p. 135

ysis

• Processing time rose with dataset size across methods (Pearson r = 0.93–0.97); on PJME and PJMW, AutoEnergy remained faster than FT and TSEff.For the two datasets exceeding 140,000 samples, AutoEnergy took 2,061s and 2,205s, respectively; evaluation on million-sample and streaming data remains future work. Across the three smallest datasets, mean nRMSE was 0.0814 and mean feature-generation time was 6.02 seconds, with lower errors and faster processing than the compared methods.Source: §4.4.2 · p. 133, ysis · p. 136

Testing

• Across datasets, AutoEnergy-generated features were the most predictive, especially significant lags and nested-window features, which capture historical effects and multiple temporal scales.Explicit cyclical, lagged, and windowed transformations yield human-readable, traceable features that support interpretation; the reported permutation-importance rankings are model-agnostic diagnostics, not complete explanations of AutoGluon’s learned forecasting functions.Source: Figure 4.8 · p. 137, Testing · pp. 138–140
• Friedman testing found significant method differences (p < .001), with AutoEnergy’s nRMSE mean rank of 1.06; Wilcoxon testing found a significant AutoEnergy–FT difference (p = .001).Bonferroni-adjusted post hoc testing found no significant AutoEnergy–FT difference (p = .820), but significant differences against TSEff, TSMin, and No.Feat. (all p < .001). For processing time, AutoEnergy ranked first (1.33); its differences from FT and TSEff were significant, but not from TSMin.Source: Testing · pp. 140–141, Figure 4.9 · p. 140, Table 4.6 · p. 141

4.5 Summary

• AutoEnergy achieved mean nRMSE 0.0338, with error reductions of 19.52% vs FT, 82.38% vs TSMin, 77.98% vs TSEff, and 84.72% vs baseline.Its mean feature-processing time was 471.94 seconds: 1.31x faster than FT and 4.41x faster than TSEff, but 1.58x slower than TSMin. The findings attribute gains to domain-specific features that capture recurring consumption patterns and improve general-purpose AutoML for ECF.Source: §4.4.2 · pp. 132–133, §4.4.1 · p. 132, §4.5 · p. 142

5.1 Introduction

• For BESS optimization, the chapter evaluates whether automated feature engineering improves downstream decisions, using a real-world UK property dataset and regret rather than forecasting error alone.The proposed end-to-end DFL framework jointly forecasts demand and electricity prices over a 24-hour horizon while optimizing battery operations; it aims to reduce reliance on domain expertise and assess practical viability against PTO and DFL approaches with and without AFE.Source: ENERGY STORAGE OPTIMISATION · pp. 144–145, §5.1 · p. 146

5.2.1 Problem Definition

• The BESS problem forecasts unknown electricity prices and household demand for an optimization model that schedules battery charging and discharging while minimizing costs and satisfying demand and operating constraints.The proposed framework uses AFE to enrich the data with domain-specific features and support decision quality in DFL.Source: §5.2.1 · p. 147

5.2.1.1 Multiple Unknowns, Decision-Making, and Objectives

• The BESS task forecasts electricity price and demand over a 24-hour horizon, then selects charging and discharging amounts while minimising total electricity cost.The objectives also include meeting household demand and respecting battery power, capacity, and state-of-charge constraints.Source: §5.2.1.1 · p. 149

age System Optimisation Problem

• The BESS scheduling problem is formulated as a MILP that minimises electricity cost while enforcing demand satisfaction, energy-flow, charge/discharge, and state-of-charge constraints.The optimisation is solved jointly over the full planning horizon, coupling decisions across intervals.Source: age System Optimisation Problem · pp. 150, 152, 153

5.2.4 Automated Feature Engineering

• AutoEnergy generates temporal features from timestamps and lag and nested-window statistics from historical consumption, including cyclical sine/cosine transformations and statistically significant lags.The method uses Kendall’s tau to select meaningful temporal relationships and aims to reduce reliance on domain expertise for feature engineering.Source: §5.2.4 · pp. 154–155

5.3.1 Datasets

• The experiment used 55 days of hourly real-world UK electricity demand and BESS data, with about 27 days for initial training, 14 for validation, and 14 for testing.ANN-based PTO, SPO+, and DBB models forecast next-day price and demand; training and validation data were later combined to fit the final models.Source: §5.3.1 · p. 157, §5.3.3 · pp. 159–160

5.3.4 Evaluation Metrics and Statistical Tests

• Decision quality was assessed with normalised regret, and non-parametric tests compared methods and AFE conditions using a 0.05 significance threshold.The evaluation plan specifies the Friedman Aligned-ranks test for overall differences and Wilcoxon Signed-Rank tests for pairwise comparisons.Source: §5.3.4 · pp. 163–164

Compared to Decision-Focused Learning

• SPO+ with AFE achieved average regret of 0.0672 versus 0.2046 for PTO with AFE, a reduction of approximately 67.16% across the fourteen-day test horizon.The results attribute the lower-cost schedules to training with SPO+ loss that accounts for downstream optimisation rather than only pointwise forecast error.Source: Compared to Decision-Focused Learning · p. 165
• DBB achieved the highest average regret, and the study suggests it may be less suitable for BESS optimisation with scarce data.The passage proposes that DBB’s design for linear-objective problems and its need for data to learn gradient approximations may contribute to its performance.Source: Compared to Decision-Focused Learning · p. 165
• SPO+ with AFE had a lower standard deviation of regret (0.0780) than PTO (0.1732) and DBB (0.2128), indicating more consistent day-to-day performance in this test.The passage notes that other approaches were more sensitive to volatile market conditions and qualifies the finding as data-dependent.Source: Compared to Decision-Focused Learning · p. 166

ing Methods with Feature Importance Analysis

• 56.48%, 19.42%, and 22.87%: AFE reduced mean regret for SPO+, PTO, and DBB, respectively, while SPO+ (AFE) achieved the lowest mean regret, 0.0672.Regret measures additional electricity cost in GBP; compared with PTO (AFE), SPO+ (AFE) reduced mean regret by 0.1374, or 67.16%. The difference between the highest and lowest mean regrets, DBB without AFE (£0.3187) and SPO+ with AFE (£0.0672), equates to approximately £91.80 per year if sustained.Source: Figure 5.6 · p. 167, ing Methods with Feature Importance… · pp. 167–169, Table 5.5 · p. 168

Methods

• AFE feature importance varied by test day, while nested rolling-window statistics and lag features consistently ranked above temporal features across PTO, SPO+, and DBB.Raw weather features were more important on some days, whereas engineered features dominated on others; the analysis suggests timestamp and historical-consumption features can improve predictions when exogenous data are unavailable.Source: Figure 5.7 · p. 171, Methods · pp. 174–175

ing Methods with Feature Importance Analysis

• The end-to-end setup favours predictors that support gradient-based training and uses point forecasts, so it excludes some high-performing regressors and does not propagate forecast uncertainty into risk-aware optimisation.A sequential PTO alternative could use other regressors, but would weaken end-to-end decision alignment and increase exposure to objective misalignment and error propagation.Source: ing Methods with Feature Importance… · pp. 173–174

5.4.4 Statistical Significance Analysis

• On the fourteen-day test horizon, Friedman’s test found a global difference among six methods (p < .001); Bonferroni-corrected Wilcoxon tests found SPO+ (AFE) beat both DBB variants and PTO without AFE (adjusted p < .001), but not PTO (AFE) or its no-AFE baseline.Source: §5.4.4 · pp. 175–177, Table 5.6 · p. 176

5.5 Summary

• The end-to-end framework combines AFE with joint next-day price and demand forecasting and regret-based BESS optimisation, evaluated on a novel real-world UK property dataset.The evidence is a single case study; broader validation across sites, seasons, and operating conditions is needed before generalising the economic benefits.Source: §5.5 · pp. 178–180, streamlining ML model development by… · p. 184

RO1: To establish baseline ML models for ECF and investigate

• Validation across two FDCS sites suggests potential transferability, but colder-month data and limited site coverage leave warmer conditions and broader operational variation untested.The datasets were predominantly collected from late October to late January; facility scale, equipment, control policies, geography, and climate may affect performance and feature relevance.Source: RO1: To establish baseline ML models… · p. 186

RO2: To develop an AFE method for ECF problems, thereby

• AutoEnergy is tailored to energy time series, tested at scales up to approximately 145k samples, and evaluated against five pipelines and eighteen datasets, but larger-scale and streaming performance remains unverified.Its one-third sequence-length constraint is heuristic, interpretability evidence is limited to permutation importance, and the benchmark comparison cannot cover all rapidly evolving feature-engineering methods.Source: RO2: To develop an AFE method for ECF… · pp. 186–187
• The DFL-AFE case study used 55 winter days from one UK property, deterministic parameters, and an aggregated BESS model, limiting generalisation of its economic findings.Results may vary with geography, tariffs, markets, seasons, assets, and operational constraints; the framework uses point forecasts rather than calibrated uncertainty estimates.Source: RO2: To develop an AFE method for ECF… · pp. 188–189

6.3.1 Data Collection and Seasonal Generalisation

• Future FDCS research should collect data in warmer seasons and test models across more locations, operational scales, and regulatory contexts.This would address seasonal and geographic gaps in the current evidence base.Source: §6.3.1 · p. 189

6.3.2 Algorithm Enhancement and Scalability

• Future work could adapt AutoEnergy’s window length to temporal patterns, combine it with other feature generators or anomaly signals, and extend DFL-AFE to renewables, longer horizons, and broader objectives.Suggested directions also include ensemble, transfer-learning, reinforcement-learning, meta-learning, and alternative DFL methods, as well as balancing cost with emissions or user comfort.Source: §6.3.2 · pp. 189–190, tems · p. 190, tal Considerations · pp. 190–191

Consumption Forecasting

• AutoEnergy may transfer to energy-like series and other usage-pattern forecasting tasks, but its benefits are conditional and require validation in each target domain.Recurring cycles, autocorrelation, and multi-scale variation make its temporal features plausible starting points; non-temporal drivers, regime shifts, or rich exogenous signals may reduce their value.Source: Consumption Forecasting · pp. 193–194

6.5. REFLECTION IN LIGHT OF RECENT ADVANCES IN AI

• The thesis argues that domain-aware temporal features can strengthen data-scarce forecasting and decision quality, while emphasizing traceability and decision-level evaluation alongside newer AI methods.The reflection cautions that large-scale and streaming performance still needs testing and recommends combining emerging models with domain constraints and safeguards.Source: in AI · p. 196, §6.5 · p. 197

with AutoML

• The preliminary AutoML study improved performance across four frameworks and established an initial approach to reducing expert-dependent feature engineering for energy consumption forecasting.The thesis reflection argues that interpretable temporal features can help general learners, especially with limited data, but AutoEnergy remains heuristic and may add less value when strong exogenous drivers dominate.Source: in AI · p. 196, §6.5 · p. 197, with AutoML · p. 199

A. EXPLORING AUTOMATED FEATURE ENGINEERING FOR ENERGY CONSUMPTION FORECASTING WITH AUTOML

• Energy forecasting features commonly include historical consumption and weather, while time-based, cyclical, lag, and rolling-window features can represent temporal patterns.Energy use exhibits linear and nonlinear behavior, irregular usage patterns, and weather influences.Source: A. EXPLORING AUTOMATED FEATURE… · p. 200
• Manual feature extraction requires time, energy-domain expertise, and data-science skills to identify, validate, and extract relevant features.Energy AutoML methods can automate preprocessing, feature selection, model selection, and hyperparameter tuning, but do not generate new useful, interpretable features.Source: A. EXPLORING AUTOMATED FEATURE… · pp. 200–201
• Existing automated feature-engineering tools still require expertise to choose lags and transformations, and their energy-sector use remains limited.PCA, wavelet decomposition, and autoencoders can produce features that are difficult to interpret, a concern for understanding energy-use drivers.Source: A. EXPLORING AUTOMATED FEATURE… · p. 201
• The proposed AFE aims to improve AutoML forecasting, reduce reliance on domain expertise, support diverse energy patterns and dataset sizes, and generate explainable features.The evaluation uses eleven publicly available real-world datasets spanning different energy domains.Source: A. EXPLORING AUTOMATED FEATURE… · p. 202

Method

• The method section introduces the problem definition and then describes the proposed approach.Source: Method · p. 202

A.2.1 Problem Definition

• The method targets an energy time-series dataset of timestamp-target pairs and seeks comprehensive automatic feature engineering to improve predictive-model performance within AutoML.The problem definition represents each instance as (t_i, y_i).Source: §A.2.1 · p. 202, A. EXPLORING AUTOMATED FEATURE… · p. 203

A.2.2 Proposed Method

• The method automates extraction and selection of features from timestamps and target values, especially lag and rolling-window features.The resulting feature subsets are concatenated into a comprehensive feature vector for each instance.Source: §A.2.2 · p. 203, A. EXPLORING AUTOMATED FEATURE… · p. 205
• Time-based features represent hour, day of week, and weekday-versus-weekend patterns, while sine and cosine transformations capture repeating daily or weekly cycles.These features use temporal information to represent consumption patterns associated with time intervals.Source: §A.2.2 · p. 203
• The method looks back up to one week and uses Pearson correlation to select the top 10 lags most correlated with the target, avoiding unnecessary feature growth and supporting efficiency and generalisability.The one-week extent reflects daily and weekly patterns in energy systems, while the passage notes that selecting optimal lags can require domain knowledge.Source: §A.2.2 · p. 204
• Rolling-window maximum, minimum, mean, kurtosis, skewness, and standard deviation capture short-term trends and variability, using each selected lag as the window size.The window size follows the number of optimal lags identified in the preceding step.Source: §A.2.2 · p. 204

A. EXPLORING AUTOMATED FEATURE ENGINEERING FOR ENERGY CONSUMPTION FORECASTING WITH AUTOML

• The engineered feature vector is formed by concatenating feature subsets, and the predictive model is trained on these vectors and their target values in AutoML.The algorithm applies time-based, cyclical, lag, and statistical feature functions before training.Source: A. EXPLORING AUTOMATED FEATURE… · pp. 205–206

Notations and Definitions:

• The notation defines the dataset as timestamp-target pairs and distinguishes each extracted feature subset, the concatenated feature vector, and the AutoML-trained predictive model.The dataset table covers residential buildings, renewable energy, and regional energy consumption, with multivariate datasets containing additional features such as weather.Source: Notations and Definitions · p. 205

A. EXPLORING AUTOMATED FEATURE ENGINEERING FOR ENERGY CONSUMPTION FORECASTING WITH AUTOML

• The design calls for interpretable features and computational efficiency, and computes features only from data available at that timestamp to prevent look-ahead bias.Interpretability supports feature-importance analysis, while the efficiency requirement addresses large datasets.Source: A. EXPLORING AUTOMATED FEATURE… · p. 207

A.3 Experimental setup

• The setup covered datasets, benchmarking methods, evaluation metrics, and statistical tests.Source: §A.3 · p. 209
• An 80% training and 20% testing split preserved unseen data for performance evaluation.The exploratory split was chosen for simplicity and computational efficiency given an average of nearly 75,000 samples per dataset.Source: §A.3 · p. 209
• Data were not shuffled, preserving chronological timestamp order to retain temporal patterns in energy consumption.AutoML model selection and hyperparameter tuning were automated, so a separate validation set was not considered necessary beyond the test set.Source: §A.3 · p. 209
• The baseline scenario trained AutoML without proposed feature engineering, using univariate datasets containing only the target variable and timestamp.This minimal-input setup was intended to reflect conditions without domain-specific feature-engineering knowledge and preliminary learning with limited data.Source: §A.3 · p. 209, A. EXPLORING AUTOMATED FEATURE… · p. 210

A. EXPLORING AUTOMATED FEATURE ENGINEERING FOR ENERGY CONSUMPTION FORECASTING WITH AUTOML

• The experimental scenario trained AutoML models with the proposed feature-engineering method.Source: A. EXPLORING AUTOMATED FEATURE… · p. 210
• Both baseline and experimental AutoML settings were evaluated on the test set to compare model performance across scenarios.Source: A. EXPLORING AUTOMATED FEATURE… · p. 210

A.3.1 Datasets

• The study uses publicly available real-world datasets from related studies and Kaggle because standardised AutoML energy-forecasting benchmarks are scarce.Source: §A.3.1 · p. 210
• The eleven datasets span residential building energy consumption, renewable sources such as wind turbines, and city or regional energy consumption.Source: §A.3.1 · p. 210
• The datasets include both univariate and multivariate forecasting problems and vary in size and granularity.Source: §A.3.1 · p. 210

A.3.2 Benchmarking Methods

• H2O, TPOT, AutoGluon, and FLAML were selected as open-source AutoML benchmarks for forecasting, and all models were trained using Python 3.11.The methods were chosen to provide a benchmark across diverse state-of-the-art AutoML approaches.Source: §A.3.2 · p. 210

A. EXPLORING AUTOMATED FEATURE ENGINEERING FOR ENERGY CONSUMPTION FORECASTING WITH AUTOML

• Figure A.1 presents the proposed feature-engineering algorithm and experimental setup.Source: Figure A.1 · p. 211

A.3.3 Evaluation Metrics and Statistical Tests

• nRMSE compares each AutoML method with and without the proposed feature engineering method.It is used as the evaluation metric for forecasting error.Source: §A.3.3 · p. 211
• nRMSE provides a normalised error measure for comparing datasets with differing actual-value scales.The passage identifies this as its purpose across datasets with differing yi scales.Source: §A.3.3 · p. 211

A. EXPLORING AUTOMATED FEATURE ENGINEERING FOR ENERGY CONSUMPTION FORECASTING WITH AUTOML

• Table A.2 compares nRMSEs across datasets for AutoML methods with versus without the proposed feature engineering; lower values are better.Source: Table A.2 · p. 212
• The Wilcoxon signed-rank test assesses paired performance with versus without the proposed feature engineering.It provides statistical support for analysis of the experimental-versus-baseline results.Source: A. EXPLORING AUTOMATED FEATURE… · p. 212
• The Friedman test assesses whether AutoML methods differ significantly in performance across multiple datasets.Source: A. EXPLORING AUTOMATED FEATURE… · p. 212
• Friedman ranking orders AutoML methods by their average performance across all datasets.Source: A. EXPLORING AUTOMATED FEATURE… · p. 212

A.4.1 Overall Performance

• The proposed AFE method significantly improved nRMSE across all examined AutoML frameworks, with no cases where a baseline outperformed it.Table A.3 reports the Wilcoxon comparisons.Source: §A.4.1 · p. 213, Table A.3 · p. 214
• Across eleven datasets, AutoGluon was the top-performing AutoML framework in five instances.The datasets represent varied energy consumption environments.Source: §A.4.1 · p. 213
• Regional energy datasets showed the greatest nRMSE reductions, while wind-turbine and home-appliance datasets also improved, to a lesser degree.These results indicate applicability across regional, renewable, and residential energy domains.Source: §A.4.1 · p. 213

A.4 Results and Discussion

• The proposed method aims to automate performance-improving feature engineering while reducing the domain knowledge required for energy-consumption forecasting.The section frames its results across energy domains and AutoML methods using nRMSE, then describes Wilcoxon tests and feature-subset ablation analysis.Source: §A.4 · p. 212, §A.4.1 · p. 213

A. EXPLORING AUTOMATED FEATURE ENGINEERING FOR ENERGY CONSUMPTION FORECASTING WITH AUTOML

• The Friedman-rank figure reports N=11, P < .001, and Chi-Square = 64.636 across AutoML methods.The caption identifies the visual as average Friedman ranks for all AutoML methods.Source: Figure A.2 · p. 214

A.4.2 Ablation Analysis of Features Contribution

• AutoGluon was selected for the ablation because of its strong overall performance and average Friedman rankings in the preceding comparison.Source: §A.4.2 · p. 214
• Using all generated features produced the lowest nRMSE overall, while lags and rolling statistics beat all features on WindT and time-based/cyclical features performed worst.The comparison groups features into all features, time-based and cyclical features, and lags and rolling statistics.Source: §A.4.2 · p. 214, Table A.4 · p. 215, A. EXPLORING AUTOMATED FEATURE… · p. 215

A.5 Summary

• The AFE method reduced prediction errors across eleven datasets and improved performance across the examined AutoML frameworks.The datasets covered residential buildings, renewable energy, and regional energy consumption; the method was tested with several state-of-the-art AutoML methods.Source: §A.5 · p. 215, A. EXPLORING AUTOMATED FEATURE… · p. 216
• The exploratory AFE method was tailored to energy forecasting to improve AutoML performance and reduce the need for domain knowledge in feature engineering.It was validated on eleven datasets spanning residential buildings, renewable energy, and regional energy consumption.Source: §A.5 · p. 215

A. EXPLORING AUTOMATED FEATURE ENGINEERING FOR ENERGY CONSUMPTION FORECASTING WITH AUTOML

• Wilcoxon tests indicated statistically significant performance improvements across all examined AutoML frameworks when the proposed AFE method was used.Source: A. EXPLORING AUTOMATED FEATURE… · p. 216
• Future work will test the proposed methods on more datasets from additional energy domains, including industrial and manufacturing settings.Source: A. EXPLORING AUTOMATED FEATURE… · p. 216

• MAE measures average prediction-error magnitude without direction, while RMSE penalises larger errors more heavily and is sensitive to outliers.Source: B Complementary Results from Chapter 4 · p. 217
• MAPE expresses errors as percentages of actual values for scale-independent interpretation, but is undefined when any actual value is zero.Source: B Complementary Results from Chapter 4 · p. 217
• R2 measures the proportion of dependent-variable variance explained by a model; values closer to 1 indicate a better fit.Source: B Complementary Results from Chapter 4 · p. 217

B.1 Results Across All Test Sets with AutoGluon

• The evaluation compares feature-engineering methods across all 18 datasets using RMSE, MAE, MAPE, and R², complementing Table 4.3.Table B.1 reports RMSE and MAE, while Table B.2 reports MAPE and R².Source: §B.1 · p. 218, Table B.1 · p. 219, Table B.2 · p. 220

B.2 Results Across All Test Sets with TabPFN

• Across all 18 datasets, Tables B.3–B.4 report TabPFN results for feature-engineering methods using RMSE, MAE, MAPE, and R², complementing Table 4.4.Table B.3 covers RMSE and MAE; Table B.4 covers MAPE and R².Source: §B.2 · p. 221, Table B.3 · p. 222, Table B.4 · p. 223
• In Tables B.3–B.4, bold indicates the best result for each row and metric.Source: Table B.3 · p. 222, Table B.4 · p. 223

ated by AutoEnergy in Chapter 4

• Feature importance is estimated by shuffling a feature’s values and measuring the resulting performance drop; higher scores indicate greater importance.The output also reports the score’s standard deviation, which is unavailable when too few shuffle sets are used to estimate variance.Source: ated by AutoEnergy in Chapter 4 · p. 224

C. ANALYSIS OF THE TOP TEN FEATURES GENERATED BY AUTOENERGY IN CHAPTER 4

• Feature-importance analyses are tabulated for the Appliances, TCity, WindT, AEP, PJME, and PJMW datasets.Source: Table C.1 · p. 225, Table C.2 · p. 225, Table C.3 · p. 225
• Feature-importance analyses are also tabulated for the COMED, FE, NI, DEOK, EKPC, and Steel datasets.Source: Table C.7 · p. 226, Table C.8 · p. 226, Table C.9 · p. 227
• Further feature-importance tables cover the UNICON, Victoria, CAISO Elec, Solar Home, FDCS 1, and FDCS 2 datasets.Source: Table C.13 · p. 228, Table C.14 · p. 228, Table C.15 · p. 228

Test Days in Chapter 5

• Models were trained and evaluated across ten independent random-seed runs, with mean performance metrics and standard deviations reported to support comparisons.The passage describes this as a way to mitigate stochastic training variance and enable statistically meaningful comparisons.Source: Test Days in Chapter 5 · p. 230
• Table D.1 presents daily test regrets for Run 1.Source: Table D.1 · p. 230

D. RESULTS OF TEN INDEPENDENT RUNS ACROSS TEST DAYS IN CHAPTER 5

• Tables D.2–D.6 report daily test regrets for Runs 2–6.Source: Table D.2 · p. 231, Table D.3 · p. 231, Table D.4 · p. 232
• Tables D.7–D.10 report daily test regrets for Runs 7–10.Source: Table D.7 · p. 233, Table D.8 · p. 234, Table D.9 · p. 234

sults Across Test Days in Chapter 5

• The daily scheduling figures show the best-performing method for each test day, selected by its lowest regret across experimental runs.The passage refers to Figures E.1 and E.2.Source: sults Across Test Days in Chapter 5 · p. 236
• Green circles above zero mark charging periods, while red circles below zero mark discharging periods.Battery energy flows appear as coloured circles connected to stems.Source: sults Across Test Days in Chapter 5 · p. 236
• Orange background bars indicate direct grid-to-property energy supply when needed.Source: sults Across Test Days in Chapter 5 · p. 236
• Dashed blue lines overlay predicted electricity prices, referenced to a secondary y-axis.Source: sults Across Test Days in Chapter 5 · p. 236
• Each subplot title identifies the test-set date, best-performing method, and corresponding experimental run.Source: sults Across Test Days in Chapter 5 · p. 236
• The horizontal axis spans 24 hours (00–23), showing scheduling patterns in relation to price signals.Source: sults Across Test Days in Chapter 5 · p. 236

E. OPTIMAL BESS SCHEDULING STRATEGY RESULTS ACROSS TEST DAYS IN CHAPTER 5

• Figures E.1 and E.2 show daily BESS optimisation, with charging, discharging, direct grid supply, and predicted electricity prices encoded separately.Green circles indicate charging, red circles discharging, orange bars direct grid supply, and a dashed blue line shows predicted electricity prices on the right y-axis.Source: Figure E.1 · p. 237, Figure E.2 · p. 238