Research On Short-Term Power Load Forecasting Model And Its Improvement Method

Due to the continuous improvement of China's installed power generation capacity and the increasingly complex power load trends,facing great challenges in power system planning and operation.As an important part of power system planning and operation,the results of power load forecasting have important reference value for real-time dispatching of power systems,optimization of dispatching plans,and safe and stable operation.Therefore,research on how to improve the accuracy of load forecasting is an important issue in the direction of power systems.This paper has predicted the hourly maximum load of a place on December 1,2018,In order to contrast with the subsequent model improvement methods,a BP neural network model,a long-term,a short-term memory neural network model,and a support vector machine model were built in sequence,and certain prediction effects were achieved.The rolling forecasting method may cause the forecasting errors to accumulate over time and affect the load forecasting accuracy.In this paper,the average daily target time was divided into 6 periods,and the average value of the maximum load per hour in each period was predicted.The predicted value and the average value of the prediction result of the single load model in each period were processed to obtain the correction value in each period,and then the correction value is used to modify the hourly prediction result of the single model.The simulation results showed that the mean absolute percentage errors of the model prediction results after the correction droped by 24.76%,11.30%,and 9.03%,respectively.Different models have different advantages when applied to load forecasting.The thesis combined the three models constructed by variable weights to form a comprehensive load model.First of all,this research used each single model to predict the hourly maximum load of the week before the target day.Then,this paper input the three sets of prediction results and actual load values into the support vector machine model for training.The paper input the predicted maximum load per hour of the three models to the target day into the trained support vector machine model,and finally obtained the maximum load per hour on the final target day.The simulation shows that the mean absolute percentage error of the prediction result of the comprehensive load model was 1.14%,and the prediction accuracy of the model was significantly improved.Convolutional neural networks can mine a large amount of data hidden information,and long-and short-term neural networks have a strong ability to fit time series data.This paper built a CNN-LSTM model,constructed input information into an image input sequence,input it to the model for training,and finally predicted the maximum load per hour on the target day.The simulation results showed that the mean absolute percentage error of the model's prediction results was 1.91%,which was 20.08% lower than that of the long-term and short-term memory neural network model.