Electricity Demand Forecasting With a Modified Extreme-Learning Machine Algorithm

Abstract

To operate the power grid safely and reduce the cost of power production, power-load forecasting has become an urgent issue to be addressed. Although many power load forecasting models have been proposed, most still suffer from poor model training, limitations sensitive to outliers, and overfitting of load forecasts. The limitations of current load-forecasting methods may lead to the generation of additional operating costs for the power system, and even damage the distribution and network security of the related systems. To address this issue, a new load prediction model with mixed loss functions was proposed. The model is based on Pinball–Huber’s extreme-learning machine and whale optimization algorithm. In specific, the Pinball–Huber loss, which is insensitive to outliers and largely prevents overfitting, was proposed as the objective function for extreme-learning machine (ELM) training. Based on the Pinball–Huber ELM, the whale optimization algorithm was added to improve it. At last, the effect of the proposed hybrid loss function prediction model was verified using two real power-load datasets (Nanjing and Taixing). Experimental results confirmed that the proposed hybrid loss function load prediction model can achieve satisfactory improvements on both datasets.