Forecasting industrial electricity demand using hybrid optimization methods

Abstrak

This study presents a hybrid machine learning framework for forecasting industrial electricity consumption by comparing backpropagation neural networks (BPNN) with models enhanced through metaheuristic optimization algorithms. Using 32 years of annual data from APEC economies, the research addresses rising electricity demand driven by economic and infrastructural development. A key limitation in traditional models— underfitting due to complex data patterns—is addressed via feature selection, which identifies the most relevant variables and reduces model complexity. Five metaheuristic algorithms—cuckoo search (CS), differential evolution (DE), harmony search (HS), particle swarm optimization (PSO), and teaching–learning-based optimization (TLBO)—are applied to optimize both feature selection and BPNN training. The proposed approach improves forecasting accuracy by handling noisy inputs and capturing the nonlinear relationships common in energy datasets. Among the tested methods, TLBO consistently delivers superior accuracy and robustness across most evaluated countries. The findings contribute an effective and adaptable forecasting model with significant implications for long-term energy planning and policy development.