Explainability Spectrum Analysis of Weather Sequences in Short-Term Load Forecasting

Short-term load forecasting tasks utilize a plethora of weather sequences in order to study meaningful features that describe complex relationships between environmental parameters and the variable of load, as well as the occurrence of impactful events, thus reinforcing the predictive power of estimators. It is evident that the inclusion of those additional features increases the input dimensions and decreases the overall explainability of forecasting models, since not all of the extracted features are equally compatible with the assumptions of the regressors, resulting in models where the justification of predictive potency, the connection between the variance of the target variable and the variance of predictors, as well as the explanation of the computational flow become increasingly difficult due to ambiguity. Therefore, this analysis investigates the representativeness of weather variables and examines the explainability of interpretable short-term load forecasting structures that utilize environmental sequences towards the selection of important feature subsets that maximize the explained variance of load time series as well as the seasonal, trend and residual load components during training. This methodology considers linear regression and prominent tree-based estimators coupled with model-dependent feature importance frameworks utilizing game theoretic, permutative and impurity-based approaches for the derivation of optimal feature sets and feature set combinations in terms of explained variance scores. Additionally, two structural approaches of explainable feature set generation from the joint feature importance scores through the utilization of hierarchical and majority inclusion are compared, their execution behavior towards explainability maximization is examined, and the main observations from each stage of this analysis are discussed.