Short-Term Power Load Forecasting Based on EMGCSO-Elman

INTRODUCTION: In short-term electrical load forecasting, Elman neural architectures often suffer from convergence instability and limited prediction accuracy due to random parameter initialization. To address this issue, an Elite Multi-mode Guided Chicken Swarm Optimization (EMGCSO) methodology is developed to determine the optimal weights and biases for the Elman network. OBJECTIVES: The primary objective is to develop an EMGCSO-Elman integrated predictive framework for short-term power load forecasting. This involves designing an enhanced chicken swarm optimization algorithm that effectively guides the search process by classifying roosters based on their fitness scores into elite, failed, and ordinary categories, and leveraging optimal, random, and neighborhood optimal solution information to formulate three distinct search strategies. METHODS: EMGCSO introduces three search strategies focusing on exploitation, exploration, and a balance between the two. These strategies are guided by elite individuals, random search, and neighborhood optima, respectively. The proposed algorithm is first validated through benchmark function tests. Subsequently, it is applied to optimize the initial parameters (weights and biases) of the Elman network, thereby constructing the EMGCSO-Elman forecasting framework. RESULTS: Benchmark function test results demonstrate that EMGCSO achieves fast convergence and strong global search capability. When deployed for forecasting electric load in the short term, the EMGCSO-Elman model significantly outperforms BP, Elman, and CSO-Elman models in predictive accuracy together with stability. CONCLUSION: The proposed EMGCSO-Elman model effectively overcomes the sensitivity of Elman neural networks to initial parameter settings, demonstrating clear superiority for short-term power load prediction tasks.