Structural Ensemble Regression for Cluster-Based Aggregate Electricity Load Forecasting
Research project aiming at the design and implementation of a peak and non-peak influenced member selection strategy for the development of structurally diverse ensemble learning models on the machine learning task of short-term load forecasting, utilizing the cluster-based aggregare framework. The project presents a case study denoting the performance benefits of this strategy for high resolution load predictions over the utilization of widely-used base estimators. This work examines the behavior of adaptive stacking and voting regressors while processing clustered time series from distinct yet anonymous client groups in order to derive optimized predictions of the partial demand values, leading to an improvement to the total demand forecast. The performance evaluation of estimators on peak and non-peak observations derived from perspective sets of indices that belong to the original as well as the predicted time series forms membership sets that subsequently construct a new feature set based on the selected estimator output. As a result, this combinatorial approach achieves reduced error metrics when different cluster evaluation methods are applied such as the elbow and silhouette analysis.