Power Load Forecasting Based On Combination Model Of Support Vector Machines And LSTM Neural Network

With economic growth and rising consumer demand for electricity and distributed energy,the prominence of electric load forecasting in power system management continues to grow.Precise load predictions can minimize generation expenses,enhance economic efficiency,and guarantee the power system’s secure and stable functioning.Nevertheless,current traditional forecasting algorithms have a somewhat limited scope of application,and factors such as load data characteristics(cyclical,non-linear,nonsmooth,heteroskedastic)and data volume significantly impact these algorithms.Consequently,there is a need for improved prediction accuracy.In order to tackle these challenges,the use of support vector machine regression(SVR),known for its robust generalization capabilities,and long short-term memory(LSTM)algorithms,recognized for their powerful non-linear fitting,may offer valuable technical resources for this research.Combining the generalisation capability of SVR and the advantages of LSTM non-linear fitting,this paper designs a combined electricity load forecasting model based on SVR and LSTM algorithms and uses it for forecasting electricity load data in New South Wales from 2020 to 2021.This paper’s principal research efforts and conclusions encompass the following:(1)Preliminary analysis of the target problem.The load data is firstly preprocessed using relevant statistical methods,and after processing the data the load data is used as input features for the load forecasting model.This is followed by a brief analysis of the characteristics,influencing factors,forecasting principles and properties of the electrical loads.(2)Construction of SVR and LSTM models.SVR model development: The RBF kernel function was employed as the SVR’s kernel function;To tackle the issue of SVR model prediction results being influenced by superparameter values,a genetic algorithm optimized the regularization coefficients C and gamma coefficients.As a result,the optimized parameters were employed to create the SVR load forecasting model.The LSTM model was developed by defining the LSTM neural network’s architecture and building the LSTM power load prediction model utilizing the Tensor Flow framework.(3)Construction of a combined prediction model based on SVR and LSTM algorithm.The forecasting results of the LSTM and the optimised SVR were compared and analysed by using the SVR and LSTM algorithms to forecast the multi-factor data of the NSW region.It was found that the LSTM forecasts were more accurate for some observations and the SVR forecasts were more accurate for some observations when forecasting the same data.Under specific circumstances,the two approaches can complement one another,thereby enhancing prediction accuracy.Building on this study,a hybrid prediction model incorporating both SVR and LSTM algorithms was developed,with the final power load forecast combining the two prediction results weighted accordingly.(4)Validation of the combined SVR-LSTM prediction algorithm.The validation process,using the same dataset,compared the prediction outcomes of each method(unoptimized SVR,GA-SVR,LSTM).The validation findings indicate that the combined SVR-LSTM model more accurately reflects fluctuations in electric load values and offers better prediction accuracy than standalone SVR and LSTM models.The prediction error indicators MAPE,RMSE,and MAE are also reduced.The ultimate prediction error for the combined prediction algorithm is within 2%,demonstrating the effectiveness of the power load forecasting model.The electric load forecasting model designed in this study can be adapted for use in China’s electric load forecasting,supplying a reference point for urban electric load forecasting in the country.