Electrical Load Forecasting Model Using Hybrid LSTM Neural Networks with Online Correction

Abstract

Accurate electrical load forecasting is of great importance for the efficient operation and control of modern power systems. In this work, a hybrid long short-term memory (LSTM)-based model with online correction is developed for day-ahead electrical load forecasting. Firstly, four types of features are extracted from the original electrical load dataset, including the historical time series, time index features, historical statistical features, and similarity features. Then, a hybrid LSTM-based electrical load forecasting model is designed, where an LSTM neural network block and a fully-connected neural network block are integrated that can model both temporal features (historical time series) and non-temporal features (the rest features). A gradient regularization-based offline training algorithm and an output layer parameter fine-tuning-based online model correction method are developed to enhance the model's capabilities to defend against disturbance and adapt to the latest load data distribution, thus improving the forecasting accuracy. At last, extensive experiments are carried out to validate the effectiveness of the proposed electrical load forecasting strategy with superior accuracy compared with commonly used forecasting models.