Research On Short-term Power Load Forecasting Based On Similar Day Extraction And Deep Learnin

Short-term load forecasting is an effective means to realize modern energy efficient dispatch and solve the problem of power supply and demand,which not only provides effective data support for power supply planning in the power market,but also escorts the safe and stable operation of the power system.In this thesis,through the analysis of current domestic and foreign electric load similar day extraction theory and short-term load forecasting methods,we select the clustering analysis,decision tree technology combined with hybrid deep learning neural network to achieve accurate forecasting of short-term load demand,and design a human-computer interaction forecasting platform in parallel with the database.First,an improved algorithm PSO-KFCM is applied to cluster the known historical load curves and mine the typical electricity consumption characteristics of each class of load in a large amount of data.The improved algorithm utilizes kernel function mapping and particle swarm optimization algorithm PSO to optimize the traditional FCM clustering algorithm twice,which solves the problems that the traditional FCM algorithm is sensitive to noise and easily limited to the initial cluster center.This improved clustering algorithm is compared with five other traditional clustering algorithms,and its superiority is demonstrated by three excellent internal clustering evaluation metrics.Then,the CART decision tree for historical date load classification is built utilizing the load type as the leaf node and the multidimensional daily feature vector as the classification basis.The daily feature vector includes multiple average weather influences and date types for each day.The daily feature vector of the day to be predicted is input to the established decision tree to obtain the load type to which the predicted day belongs,and the date of the same load type is a similar day.Finally,the proposed hybrid forecasting model CNN-GRU is carried out to forecast the future 24-hour electric load data,where the CNN-GRU model not only has the ability to extract local features but also has the ability to learn long-term dependency relationships.Training data of model are load data,weather data,and date type of similar day,which avoids the risk that the nearest neighboring dates do not exhibit different electricity consumption characteristics from the day to be predicted,and enhances the prediction accuracy and training efficiency.The model proposed in this thesis is compared with the model without similarity extraction theory and other five deep learning models,and the feasibility and effectiveness of the model are testified from the multifaceted error analysis.