Residential electricity prediction based on GA-LSTM modeling

Abstract

With residential electricity demand increasing year after year due to population growth and infrastructure expansion, accurate electricity load forecasting is critical to the operation of power systems. Although certain traditional machine learning models have achieved better results in several fields, their prediction accuracy for energy sources such as electricity is still unsatisfactory. The nonlinear and non-smooth consumption patterns of residential electricity make accurate prediction of electricity loads even more challenging. In this paper, we propose an attention-based competitive stochastic search method for training Long Short-Term Memory (LSTM) networks to predict electricity energy consumption. On the one hand, an attention layer is added to the traditional LSTM to assign feature weights; on the other hand, a genetic algorithm is introduced for stochastic search to prevent convergence to a local optimum during the training process. The experimental results show that the proposed model combines genetic algorithm and LSTM, and the prediction accuracy of the proposed model is higher than other existing prediction models in terms of root mean square error (RMSE), mean absolute error (MAE), r-square(R2), and mean absolute percentage error (MAPE), which are 0.6056, 0.3726, 0.8517, and 0.1856, respectively. In addition, the model also achieves excellent results in short-term prediction, which is better than the traditional LSTM, and the accuracy is significantly improved. It is proved that this method is effective and feasible in the prediction of power energy consumption.