A Model-Agnostic Correction Framework for Short-Term Building Load Forecasting

1. Introduction

• Short-term building loads reflect interacting weather, occupancy, HVAC, thermal, and temporal factors that can shift operating regimes.These rapidly changing conditions make forecasts useful for near-term operational decisions such as demand response, HVAC scheduling, and peak-load management.Source: §1 · p. 1
• Jointly matching recent load states and weather context addresses the risk that either load-only or weather-only analogues may have incompatible future trajectories.The framework aggregates retrieved futures and fuses them with the base forecast as a correction, rather than relying on analogues as standalone forecasts.Source: §1 · pp. 2–3
• WeL-RAC is a model-agnostic retrieval-augmented correction framework that enhances existing forecasters without modifying their architectures.It uses a dual-view retrieval mechanism without additional retrieval-module training.Source: §1 · p. 3
• Retrieved cases make each correction inspectable, allowing the source and direction of beneficial or harmful adjustments to be traced.The paper includes success and failure case studies.Source: §1 · p. 3
• The study conducts experiments across 31 building cases, three datasets, eight forecasting backbones, and four forecasting horizons, with ablation, sensitivity, and computational-cost analyses.Source: §1 · p. 3

2.1. Time Series Forecasting

• Parametric forecasting models generally encode training knowledge in model parameters, which can leave sparse historical operating regimes inaccessible during inference.Retrieved historical segments offer a complementary non-parametric mechanism.Source: §2.1 · pp. 3–4
• WeL-RAC adapts dual-view retrieval to building forecasting by matching compact recent load-state summaries and horizon-aligned weather forecasts, then correcting a base forecaster without retraining.Its retrieval and fusion operate as an external post-prediction layer.Source: §2.1 · p. 4

2.2. Building Energy Forecasting

• Building analogues can provide empirical forecasting evidence, but prior similar-day methods are typically standalone strategies or training-data selection procedures rather than external corrections to heterogeneous forecasters.Prior work also motivates considering both load and weather when selecting analogues.Source: §2.2 · p. 4
• Figure 2 depicts separate forecasting and retrieval branches whose outputs are fused after prediction.The retrieval branch matches load state and horizon-aligned weather summaries and aggregates analogue futures.Source: Figure 2 · p. 5

3. Methodology

• WeL-RAC is a lightweight correction layer that attaches to different forecasting backbones without changing their internal architectures.A trained forecaster first generates an initial load prediction; retrieval then finds similar cases and fuses their aggregated futures with that prediction.Source: §3 · p. 5

3.1. Framework Overview

• At each forecast origin, the pipeline uses a historical load-covariate window for forecasting, a recent load window for matching, and horizon-aligned weather for retrieval.The retrieval branch fuses two similarity scores, selects top-K analogue futures, and aggregates them before correction.Source: §3.1 · p. 5

3.2. Problem Definition and Formulation

• The retrieval query combines a weather summary of the forecast horizon with a recent load-state summary, while the base model predicts one target load channel from multivariate inputs.The load-state encoder summarizes recent levels, variability, extrema, net change, trend, and short-term differences.Source: §3.2 · pp. 5–6
• The correction module fuses the initial forecast with aggregated analogue evidence, and the retrieval memory uses training-period data only.Thus, the base model supplies the initial prediction while retrieval supplies instance-specific evidence without using validation or test futures.Source: §3.2 · p. 6

3.3. Forecasting Branch

• The base forecaster maps the historical load and covariates window to an initial prediction for the retrieval branch to correct.It may be a linear temporal model, tree-based regressor, neural sequence model, Transformer, or pretrained time-series foundation model.Source: §3.3 · p. 6
• WeL-RAC imposes no architectural constraint on the base forecaster and does not modify its training objective.The retrieval module operates as an external correction layer decoupled from the base model’s internal representation.Source: §3.3 · p. 6
• The forecasting branch’s initial prediction is the output that the correction framework subsequently adjusts.The forecasting and retrieval branches therefore have distinct roles in the framework.Source: §3.3 · p. 6

3.4. Weather Summary Construction

• The weather-summary channel applies the same deterministic operator to current forecast and historical horizon-aligned weather windows.Using the same template and field-selection rules makes their weather contexts comparable in textual space.Source: §3.4 · p. 7
• Weather-summary fields are conditional: temperature and time are used when available, while moisture, precipitation, and wind clauses depend on dataset variables.A frozen text encoder maps each summary to an L2-normalised embedding for semantic retrieval over weather regimes.Source: §3.4 · p. 7

3.5. Load-State Summary

• The load-state branch converts the recent load-only sequence into a compact numerical summary for similarity matching, not direct future-load forecasting.In the default setting, the encoder maps the recent segment to a 9-dimensional statistical summary.Source: §3.5 · p. 8
• The load-state summary combines demand level, central tendency, dispersion, range, net change, long-window trend, and short-window ramping behaviour.It is standardised using training-set statistics and L2-normalised before cosine similarity; the pure-load-sequence ablation instead uses the normalised recent window itself.Source: §3.5 · p. 8

3.6. Retrieval Memory and Similarity Fusion

• Each horizon-specific memory item pairs a historical load-state summary and horizon-aligned weather-summary embedding with its realised future load trajectory.Only training-set forecast origins with an observed full future horizon enter the memory.Source: §3.6 · p. 8
• The analogue score combines cosine similarities for the weather summary and load-state summary, with λ controlling their relative contribution.λ=0 gives weather-only retrieval, λ=1 gives load-state-only retrieval, and intermediate values match both weather relevance and operational compatibility.Source: §3.6 · p. 8

3.7. Retrieval of Analogue Futures

• The retrieved future evidence is a temperature-scaled softmax-weighted average of the realised futures associated with the selected top-K analogues.The temperature parameter τ controls the concentration of neighbour weights.Source: §3.7 · pp. 8–9
• The resulting trajectory is a non-parametric analogue forecast induced from historical futures resembling the current load state and weather regime.The study evaluates cases drawn from AlphaBuilding, Arizona State University, and University of Central Florida.Source: §3.7 · p. 9

• Retrieved future load trajectories provide an evidence-based correction target for the initial prediction.Source: §3.8 · p. 9
• Equation (19) instantiates the correction module as a lightweight weighted fusion.Source: §3.8 · p. 9
• At α = 0, WeL-RAC reduces to the base forecaster, while at α = 1, the output relies entirely on retrieved analogue futures.Source: §3.8 · p. 9
• Intermediate α values provide conservative correction when a strong base model can still benefit from similar weather-load evidence.Source: §3.8 · p. 9

4. Experimental Setup

• The setup compares base forecasts with retrieval-corrected forecasts using the same chronological split, horizons, and tuning procedure.Source: §4 · p. 9

4.1. Datasets and Preprocessing

• Three benchmark datasets use multivariate hourly building-load series aligned with available building covariates and weather information.Source: §4.1 · p. 9

4.1.1. Synthetic Dataset

• AlphaBuilding contributes nine simulated two-year load–weather series spanning three cities and three energy-efficiency configurations.Source: §4.1.1 · p. 9

4.1.2. Real-world Datasets

• The real-world benchmarks comprise 12 ASU buildings and 10 quality-filtered UCF buildings, with their respective electricity-demand targets and auxiliary data.ASU series cover 2017–2019 and contain 26,280 hourly records; UCF series contain 26,278 hourly records.Source: §4.1.2 · p. 10

4.1.3. Forecast Weather Inputs

• AlphaBuilding uses oracle future weather, while ASU and UCF use site-level HRRR forecasts with lead times from 1 to 18 hours.Forecast origins without complete weather coverage are excluded for all methods; at H = 18, this removes 48 ASU and 180 UCF origins.Source: §4.1.3 · p. 10

4.2. Baselines and Metrics

• Eight diverse forecasting backbones are evaluated, and the reported metrics are MSE, MAE, R2, and MAPE.Each backbone is paired with its retrieval-corrected prediction to measure correction without changing the forecaster architecture.Source: §4.2 · pp. 10–11

4.3. Forecasting Protocol

• Each building is split chronologically 7/1/2 into training, validation, and test periods, with retrieval memory restricted to training data.The shared evaluation uses a 24-hour input window; Chronos-2 remains pretrained and zero-shot.Source: §4.3 · p. 11

4.4. Hyperparameter Search

• Forecaster and retrieval-correction hyperparameters are selected separately on validation data, with the held-out test period reserved for final reporting.Table 4 presents search spaces for the eight forecasting models.Source: §4.4 · p. 11, Table 4 · p. 12

5.1. Overall Forecasting Performance

• 90 of 96 comparisons favor retrieval-enhanced forecasts over their corresponding base forecasters across the three datasets and eight backbones.MSE, MAE, and R^2 improve for every backbone–dataset combination; MAPE is the main exception, particularly on UCF.Source: §5.1 · p. 13, Table 6 · p. 14

5.2. Building-wise Results

• 30 of 31 buildings improve in MSE and 28 of 31 in MAE, with gains most consistent on AlphaBuilding.All AlphaBuilding buildings benefit, with average reductions of approximately 34.0% in MSE and 24.6% in MAE; UCF is more heterogeneous.Source: §5.2 · p. 13, Figure 4 · p. 15

5.3. Horizon-wise Results

• Across AlphaBuilding horizons, MSE improvements remain above 32% and MAE improvements above 22%; on ASU and UCF, MSE gains grow at longer horizons.ASU MSE improvement changes from −2.9% at H=1 to 10.3% at H=18; UCF changes from 2.8% to 9.4%.Source: §5.3 · pp. 13–14, Figure 5 · p. 15

5.4. Comparison with Existing Retrieval-Based Methods

• Load-only WeL-RAC achieves lower MSE and MAE than all three retrieval baselines on every dataset, while the full model wins 10 of 12 dataset–metric comparisons.Adding the weather branch further reduces MSE relative to load-only WeL-RAC on all three datasets.Source: §5.4 · pp. 14–16, Table 7 · p. 16

5.5. Ablation Studies

• The full ablation variant is best in 7 of 12 dataset–metric comparisons; load-state similarity is the primary retrieval signal, with structured weather contributing complementary information.Replacing the weather summary with direct numeric weather inputs consistently degrades performance across all three datasets.Source: Table 8 · p. 16, §5.5 · pp. 16–17

5.6. Sensitivity Analysis

• MSE is relatively insensitive to λ and K, nearly unchanged across tested τ values, and most sensitive to α, with minima around 0.3 at H=1 and 0.4 at longer horizons.The sensitivity analysis uses DLinear and PatchTST on ASU Building 10 across four horizons.Source: Figure 6 · p. 17, §5.6 · pp. 17–18

5.5. Ablation Studies

• Load summaries outperform raw-sequence matching on ASU and UCF, whereas raw sequences work better on AlphaBuilding’s repetitive patterns.The proposed state representation compresses each retrieval item from 24 dimensions to a nine-dimensional vector.Source: §5.5 · p. 17

5.7. Computational Cost

• On UCF Building 30 at H=18, retrieval memory takes 5.668 s to build and uses 28.775 MiB; retrieval and fusion take 6.135 ms per sample.Post-hoc fusion takes 7.232 s for the 5,238-sample test set.Source: Table 9 · p. 18, §5.7 · p. 18

6.1. Successful case studies

• MSE fell from 81.33 to 18.86 (76.81%) and from 961.79 to 481.19 (49.97%) in two selected success cases, as retrieval corrections opposed MLP errors.Both used the fixed validation-selected fusion weight α=0.2; the examples do not represent average test-set performance.Source: §6.1 · p. 18, Figure 7 · p. 19

6.2. Failure case studies

• Across three seeds, retrieval fusion raised test MSE from 197.03 to 228.32 (15.88%) for the UCF Building 30 MLP at H=18 h.The two plotted windows were selected post hoc to diagnose failure, with the validation-selected 24-hour input, Top-10 retrieval, and α=0.5 configuration retained.Source: §6.2 · pp. 18–19, Figure 8 · p. 20
• MSE rose from 1455.97 to 7333.66 when retrieval favored recovery after an abrupt drop to a sustained low-load episode absent from validation.The retrieved weighted mean was 150.06 while the observed future mean was 17.60; the fixed fusion weight of 0.5 raised the final mean prediction to 102.31.Source: §6.2 · p. 19, Figure 8 · p. 20
• For a stable low-load query, similar temporal shapes still had incompatible scales: retrieved futures averaged 86.98 kWh versus 17.82 kWh observed.Retrieval-weighted load-state and weather cosine similarities were 0.980 and 0.988, yet fusion shifted the MLP mean from 17.18 to 52.08 kWh and raised MSE from 0.96 to 1181.92.Source: §6.2 · pp. 19–20

6.3. Discussion and Implications

• Retrieved trajectories make each correction’s source and direction inspectable, while failure cases expose mismatches such as unseen conditions or incompatible absolute load levels.Successful corrections can help through error complementarity even when the retrieved future is not independently more accurate than the base forecast.Source: §6.3 · p. 20
• The discussion proposes broader-regime validation, load-level-aware neighbor selection, reliability-aware fusion, and adaptive memory updates to improve retrieval reliability.A reliability-aware weight could reduce retrieval’s influence when evidence is poorly matched or strongly inconsistent with the base prediction.Source: §6.3 · p. 20

7. Conclusion

• WeL-RAC retrieves observed historical continuations by jointly matching recent load states and horizon-aligned weather summaries, then fuses them with initial predictions without changing the base forecaster’s architecture or training objective.Source: §7 · pp. 20–21
• Lower MSE and MAE and higher R^2 were reported for all 24 backbone-dataset combinations; after averaging across backbones and horizons, MSE and MAE improved for 30 and 28 of 31 building cases.The evaluation covered 31 building cases, eight forecasting backbones, and four prediction horizons.Source: §7 · p. 21
• Ablations identified load-state matching as the primary retrieval signal and weather summaries as complementary context, with the better-performing state representation varying across datasets.Compact state representations were more effective on ASU and UCF, while raw load sequences performed better on AlphaBuilding.Source: §7 · p. 21
• The retrieval memory and validation-selected fusion weight remain fixed during testing, and high similarity in normalized retrieval space does not guarantee compatible absolute load levels.The evaluation is limited to offline, single-building point forecasting on selected synthetic scenarios and two campus datasets.Source: §7 · p. 21