Short-term electric load forecasting based on series decomposition and Meta-Informer algorithm

Abstract

Accurate electricity load forecasting is crucial for the effective development of power management strategies. However, achieving both forecasting accuracy and efficiency is often challenging. This paper presents a novel framework that integrates Seasonal-Trend decomposition using Loess (STL), clustering, and meta-learning for electricity load forecasting. First, the local regression-based STL method decomposes the electricity load data into trend, seasonal, and residual components. Next, data slicing and clustering are performed based on seasonal and residual patterns. Using the Local Search k-means++ with Foresight(FLS++) clustering method, we expand the clustered data to generate multiple training tasks, which are then trained using the meta-learning-based Meta-Informer forecasting model. Subsequently, we assess the smoothness of the seasonal and residual testing tasks using the Standard Differenced Smoothness (SDS) metric. Adaptive filtering processes the data, and the model is fine-tuned for accurate predictions. Additionally, we employ a BiGRU model to forecast the trend component, which is summed and reconstructed to yield the final prediction results. Experimental results demonstrate that our approach effectively enhances forecasting accuracy in electricity load prediction.