Short-term electrical load curve forecasting with MEWMA-CP monitoring techniques

Abstract

Load forecasting is an essential component in the power sector for effective demand-side management. A decline in forecasting accuracy can significantly compromise the efficacy of planning and management strategies, particularly in the face of substantial changes to the underlying model structure. To mitigate this challenge, rigorous model monitoring is imperative to ensure the electrical system${}^{{}^{\prime }}$s reliable operation. Based on radial basis function neural network (RBF-NN) and least square support vector machine regression (LS-SVMR), an innovative prediction framework for multivariate exponential weighted moving average with cautious parameter learning (MEWMA-CP) control scheme is proposed in this paper. Central to this framework is the continuous monitoring and analysis of the residual sequence for daily electrical load data. This detailed examination allows us to meticulously track the distributions of key model features. When a significant deviation in data distribution is detected, indicating a shift from historical patterns, the proposed MEWMA-CP control scheme is activated. This scheme serves as an early warning system, triggering alerts that necessitate timely updates to the forecasting model. The MEWMA-CP control scheme is a groundbreaking addition to load forecasting methodologies, designed to ensure that the model remains current and accurate, providing a solid foundation for policy formulation and the strategic planning of future installed power capacities. The adaptability of our method to update model parameters in response to detected data distribution shifts is a distinguishing feature that sets it apart from conventional approaches. Empirical evidence from our experimental validation demonstrates the method’s capability to promptly detect changes in data distribution and dynamically update the model parameters, thereby achieving more precise and reliable prediction outcomes.