The Combination Model Of Deep Learning And The Application In Electric Power Forecasting

With the transformation of energy sustainable development,electric power is an important factor to achieve the goal of "double carbon",and the accurate prediction of electric energy consumption is particularly important for the safe and stable operation of energy supply and power grid.However,the volatility,uncertainty and instability of power consumption have brought certain challenges to its prediction.In order to solve the prediction problem of household power consumption in the power grid,this thesis studies and establishes the prediction model of power consumption.The main works are as follows.This thesis chooses the time series of historical power consumption as the feature variables.Landmark-based spectral clustering(LSC)and a deep learning model are used to cluster and predict the power consumption dataset respectively.Firstly,the experimental data is converted into a matrix,and all missing values are recovered by matrix completion.Secondly,according to the periodicity and regularity of power consumption,the data samples are divided into three clusters using LSC method.Then,the samples in each cluster are expanded by bootstrap aggregating method.Subsequently,the combined model of convolution neural network(CNN)and long short-term memory(LSTM)is adopted to predict the power consumption.CNN mainly extracts features from input data in sequence learning,while LSTM aims to train samples and predict power consumption.Finally,the prediction performance of LSC-CNN-LSTM is compared with other deep learning models to verify its reliability and effectiveness in the field of household power load forecasting.The experimental results show that the proposed hybrid method is superior to other advanced deep learning techniques in prediction performance.In order to reduce the time cost of processing large datasets,and ensure the double optimality of accuracy and time,a prediction model is proposed on the basis of the combination of variational mode decomposition(VMD)and broad learning system(BLS).Firstly,the preprocessed time series is decomposed into several modes through VMD,and multiple more stationary sub-sequences are obtained with different frequencies.The input and output features are then generated from all extracted modes.Next,the constructed samples are fed into the BLS,and the mapping features and enhancement nodes are combined to predict the global active power of household electricity.Finally,the experimental results show that the VMD-BLS model is superior to other deep learning models in prediction accuracy and time complexity,which verifies its reliability and efficiency.The hybrid prediction model of clustering algorithm and deep learning improves the accuracy of power consumption prediction.While the hybrid prediction model of signal decomposition and broad learning further improves the time cost of model training,which showing the advantages of accuracy and time in power consumption prediction.