Optimization of power system load forecasting and scheduling based on artificial neural networks

Abstract

This study seeks to enhance the accuracy and economic efficiency of power system load forecasting (PSLF) by leveraging Artificial Neural Networks. A predictive model based on a Residual Connection Bidirectional Long Short Term Memory Attention mechanism (RBiLSTM-AM) is proposed. In this model, normalized power load time series data is used as input, with the Bidirectional Long and Short Term Memory network capturing the bidirectional dependencies of the time series and the residual connections preventing gradient vanishing. Subsequently, an attention mechanism is applied to capture the influence of significant time steps, thereby improving prediction accuracy. Based on the load forecasting, a Particle Swarm Optimization (PSO) algorithm is employed to quickly determine the optimal scheduling strategy, ensuring the economic efficiency and safety of the power system. Results show that the proposed RBiLSTM-AM achieves an accuracy of 96.68%, precision of 91.56%, recall of 90.51%, and an F1-score of 91.37%, significantly outperforming other models (e.g., the Recurrent Neural Network model, which has an accuracy of 69.94%). In terms of error metrics, the RBiLSTM-AM model reduces the root mean square error to 123.70 kW, mean absolute error to 104.44 kW, and mean absolute percentage error (MAPE) to 5.62%, all of which are lower than those of other models. Economic cost analysis further demonstrates that the PSO scheduling strategy achieves significantly lower costs at most time points compared to the Genetic Algorithm (GA) and Simulated Annealing (SA) strategies, with the cost being 689.17 USD in the first hour and 2214.03 USD in the fourth hour, both lower than those of GA and SA. Therefore, the proposed RBiLSTM-AM model and PSO scheduling strategy demonstrate significant accuracy and economic benefits in PSLF, providing effective technical support for optimizing power system scheduling.