Short-term Load Forecasting Method Based On Deep Neural Network Considering Generalized Demand Side Resources

With the increasing integration of generalized demand side resources such as distributed generation and energy storage,a large number of idle adjustable resources appear on the demand side.Under the development trend of power market gradually opening,users participate in market regulation in the way of demand response under the integration of load aggregation mechanism,which enhances the load transfer capability and changes the load characteristics and variation regularity.The enhancement of randomness and flexibility of load puts forward higher requirements for the accuracy and stability of load forecasting.Therefore,the generalized demand side resources should be considered in the short-term load forecasting,and the fast and efficient load forecasting method should be selected to improve the accuracy of load forecasting.Firstly,the generalized demand side resource scheduling problem based on load aggregator(LA)is researched in this dissertation.In the real-time electricity price environment,a generalized demand side resource scheduling model based on the electricity price contract is proposed with the goal of maximizing LA revenue.The model can be used to determine the optimal scheduling plan of generalized demand side resources,guide users to reasonably arrange the time of electricity consumption,and achieve the effect of peak load shifting.Then,in this scheduling mode,a short-term load forecasting method based on deep belief network(DBN)is proposed.The optimal scheduling plan of generalized demand side resources is incorporated into the DBN load forecasting model as an additional network input,and a DBN short-term load forecasting model considering generalized demand side resources is established.The effectiveness and superiority of the improved short-term load forecasting method are verified by simulation experiment and analysis.Considering the strong temporal correlation of load,the long short-term memory(LSTM)network with good memory function was selected for load forecasting,and a short-term load forecasting model based on deep LSTM network was proposed.The model is composed of several LSTM networks and a fully connected output layer,and Adam algorithm is used to train and optimize the network.Simulation results show that compared with other methods,LSTM network can effectively process the regularity information in time series data and has higher forecasting accuracy.By comparing the forecasting results of the models with different inputs,it is shown that the forecastingerror of the model considering the generalized demand side resources is smaller,which further verifies the effectiveness of the model considering the generalized demand side resources.Finally,in order to avoid the limitation of single model,a load prediction method based on DBN-LSTM combined model based on the characteristics of DBN and LSTM networks was proposed in this dissertation.On the basis of DBN and deep LSTM network forecasting model above,a linear regression model is established to find the linear relationship between load forecasting value of single model and the true value,assign the appropriate weight to the single model and conduct a linear combination of the forecasting results.Experimental results show that the DBN-LSTM combined model can further improve the forecasting accuracy,the forecasting error is smaller than the single model,and has better stability.