Abstract:

As the basis for maintaining the operation and analysis of the power grid system, short-term power load forecasting provides judgment basis and information for the economic dispatch and safety analysis of the power grid system, and plays an important role in maintaining the normal operation of the power grid system. In this study, the FA(Firefly Algorithm) is used to optimize the penalty factor c, nuclear parameter g of SVR(Support Vector Regression) model and the number of neurons m and learning rate lr of LSTM(Long Short-Term Memory) model. The FA-SVR-LSTM combined prediction model is established using the optimal parameters, and the sample data are predicted. Taking the historical data of power load of Florida as an example, four reference models of LSTM, SVR, FA-SVR and FA-LSTM are established to predict the power load of 360 h in 15 days, and the results are compared with those of FA-SVR-LSTM. The experimental results show that compared with LSTM and SVR model, the prediction accuracy of FA-SVR-LSTM model is improved by 33.184 9% and 30.326 5%, respectively. The evaluation values of MAPE and RMSE are significantly lower than those of the other four models. These results indicate that the prediction effect of FA-SVR-LSTM combined model is significantly improved when compared with other models.

Key words: power load forecasting, prediction accuracy, firefly algorithm, long short-term memory neural network, support vector regression, combination model, parameter optimization, error evaluation