Short-term load forecasting using neural networks combined with linear models

This paper presents a short-term load forecasting approach for power systems. This approach combines neural networks with linear models. The electric load is assumed to consist of several components. Some components are described with linear models, and the others are captured with multilayer feedforward neural networks. The combination of neural networks with linear models brings advantages of them to this short-term load forecasting approach. After training with adequate policies, each component model describes the corresponding load component veritably This ensures the robustness of the forecasting approach. Testing this approach with load and weather data reveals satisfactory performance with mean absolute percentage error 3.14% for ahead-time within one day and 3.70% for ahead-time within one week.