Source-linked AI summary
On the global feature importance for interpretable and trustworthy heat demand forecasting
Milan Zdravković
TL;DR
Heat-demand forecasting supports consequential district-heating decisions, creating a need for models whose behavior stakeholders can interpret and validate. The paper evaluates global feature importance using gradient-boosting models and four interpretability approaches, finding that ALE reflects control-curve behavior while SHAP accounts for interactions and nonlinearities robustly. The paper also shows that PD can be misleading when feature effects are heterogeneous or correlated.
Problem
Heat-demand forecasting influences compliance, customer satisfaction, and liability-relevant service levels, while complex models can be difficult for stakeholders to justify and validate.
Method
The paper evaluates global feature importance using intrinsic gradient-boosting importance, Partial Dependence, Accumulated Local Effects, and SHAP without feature permutation or perturbation methods that may introduce unrealistic values.
Results
ALE reproduces behavior resembling the district-heating control curve, while SHAP accounts for feature interactions and nonlinearities and is assessed as the most reliable and robust method.
Takeaways & Limitations
Global feature-importance interpretations can support validation of heat-demand forecasting models and build trust in predictions used for district-heating operations.
Takeaways & Limitations
Partial Dependence ignores feature interactions and can be unreliable with correlated features, while averaging can obscure heterogeneous effects.
Abstract
from arXiv · showhide
The paper introduces the ante-hoc Explainable AI methodology to assess the global feature importance of the Machine Learning models used for heat demand forecasting in intelligent control of District Heating Systems, with motivation to facilitate their interpretability and trustworthiness, hence addressing the challenges related to adherence to communal standards, customer satisfaction and liability risks. Methodology includes use of four different approaches, namely intrinsic interpretability of Gradient Boosting method and selected post-hoc methods, namely Partial Dependence, Accumulated Local Effects and SHAP. None of the selected methods assume feature permutation or perturbations which can introduce bias due to introduction of random unrealistic values of data instances. Discussion of results is provided, including the assessment of complementarities where applicable, with specific interpretations in context of the district heating processes.
1 Introduction
Machine learning supports proactive, multivariate heat-demand forecasting for district heating control, but the resulting decisions can be difficult for stakeholders to justify. Explainable AI is presented as a way to improve transparency, interpretability, accountability, bias detection, and trust.
- Multivariate heat-demand forecasting can replace ambient-temperature control curves with proactive machine-learning control of District Heating Systems.
- Although heat-demand forecasting has predictable hourly patterns, its control role creates risks involving compliance, customer satisfaction, and contractual liability.
- Forecasting algorithms may be difficult for domain experts and plant operators to understand and justify because of their intrinsic complexity.
- Ante-hoc XAI uses inherently interpretable models, whereas post-hoc XAI explains typically black-box models after training and deployment.
- XAI methods can be model-agnostic or model-specific, and local or global depending on whether they explain individual predictions or entire model behavior.
- Global XAI insights can provide human-readable support for experts checking model validity, particularly in trustworthy DHS automation and out-of-distribution forecasting settings.
2 Methodology
The methodology evaluates global feature importance for a gradient-boosting heat-demand model using intrinsic model importance and three complementary interpretability methods. Historical SCADA, meteorological, and engineered time-series data are prepared for this analysis.
- The test model uses ensembles of decision trees so intrinsic and post-hoc feature-importance metrics can be compared and combined.
- The demonstration uses four heating seasons of SCADA data from DHS substation 17, merged with meteorological data to add predictors.
- Preprocessing retains full-hour observations, inserts missing timepoints, interpolates missing values, removes non-heating-season data, and filters features by correlation.
- Figure 1 presents line plots of selected data features over an example interval.
- The study compares intrinsic gradient-boosting feature importance with Partial Dependence, Accumulated Local Effects, and SHAP.
3 Overview of global feature importance insights and discussion
The paper combines intrinsic Gradient Boosting importance measures with Partial Dependence, ICE, ALE, and SHAP to interpret global feature importance in heat-demand forecasting. These approaches reveal complementary patterns linked to daily seasonality, temperature effects, feature interactions, and model decision behavior.
- Intrinsic importance measures: Gradient Boosting exposes gain, cover, and frequency as distinct global feature-importance dimensions.Gain measures accuracy improvement, cover measures the share of affected observations, and frequency measures how often a feature is used in tree splits.
- Intrinsic importance measures: Ambient temperature is used most frequently, while previous-day transmitted energy provides the largest performance gain in Gradient Boosting splits.The model contains 100 boosted trees; previous-day nearby-substation temperature also has broad coverage across test instances.
- Partial Dependence and ICE: Partial Dependence and ICE show strong hourly seasonality, nonlinear ambient-temperature effects, and heterogeneous contributions from previous-day transmitted energy.The hour_of_day effect rises from early morning through late afternoon and drops during late evening and early night; ambient temperature reverses near tamb=13.
- Partial Dependence and ICE: The high use of tamb relative to the similar temp signal is consistent with the temperature response observed in the Partial Dependence analysis.Under the stated lack-of-interaction assumption, the contrast can be interpreted through model decision points.
- Accumulated Local Effects: ALE provides a correlation-robust view in which transmitted-energy effects decrease almost linearly below tamb<9 before becoming more rapid and nonlinear.The ALE temperature distribution is described as more stable than the corresponding Partial Dependence result, while deltae-23 remains flat.
- SHAP: SHAP indicates that deltae and hour_of_day have the greatest impact on predictions, with high deltae increasing forecasts and low deltae decreasing them.The direction of the SHAP effects is reversed for features such as tretsec and feelslike, and the reported patterns correspond to district-heating behavior.
- Complementarity and caveats: Gain can be compared with SHAP to assess whether features improving training accuracy most also exert the strongest influence on predictions, whereas weight and cover are complementary measures.Intrinsic metrics use training data, while post-hoc importances use test data; frequent or broad feature use need not imply a large prediction effect.
4 Conclusions
The paper applies four global feature-importance methods to validate heat-demand forecasting models for district-heating environments requiring legal compliance and agreed service levels. The interpretations connect model behavior with district-heating processes, while highlighting methodological differences and limitations.
- Global feature-importance assessment: Four methods—ante-hoc feature importance, Partial Dependence, Accumulated Local Effects, and SHAP—are used to interpret a Gradient Boosting heat-demand forecasting model.Permutation Importance and LIME are excluded because feature permutation or perturbation can introduce unrealistic values and bias.
- Method complementarities and limitations: Partial Dependence captures only main effects and can be unreliable when features interact or are mutually correlated.ICE can expose heterogeneous effects that PD averages out, while ALE accounts for interactions and nonlinearities more reliably.
- District-heating interpretation: ALE shows transmitted energy decreasing almost linearly for tamb<9, followed by a faster, nonlinear decrease at warmer temperatures.This behavior resembles the district-heating control curve relating secondary-supply water temperature to ambient temperature.
- Method complementarities and limitations: SHAP accounts for feature interactions and model nonlinearities, making it the paper’s most reliable and robust global-importance method.The paper presents this assessment as demonstrated for the heat-demand forecasting model.
- Operational relevance: Global feature-importance interpretations can support trust in forecasting models and guide heating schedules or energy-use optimization.The intended application is operational validation in district-heating services serving many customers under compliance and service-level requirements.