Research On Short-term Load Forecasting Methods Based On Machine Learning

Accurate short-term power load forecasting helps to improve the operating efficiency of the power system and reduce operating costs.The traditional method based on statistical learning has high prediction accuracy and is easy to implement,but it performs poorly when the load fluctuation is severe.The machine learning method has a strong ability to fit,it can fully reflect the nonlinear characteristics of the load,and has achieved good results in the field of short-term load forecasting.This paper studies the short-term load forecasting method based on various methods in machine learning combined with actual data.The main works of this paper are as follows:Firstly,the construction of short-term power load feature set is studied.This part first preprocesses the data,to ensure the accuracy of subsequent model prediction,the missing values in the data set are complemented,and the outliers are eliminated.Then it is necessary to determine the attributes that affect the load change,for the large number of lag historical load characteristics,the gradient lifting tree recursive feature elimination method is used to select the features with higher importance to improve the prediction accuracy of the model and reduce the computational complexity of the model.Finally,in order to further mine the information in the historical load sequence,the deep self-encoder is used to construct new features by extracting the historical load sequence,completing the construction of feature sets to lay the foundation for subsequent predictive models.Then,the construction of short-term electric load forecasting model based on wavelet neural network is studied.The model first introduces two update strategies of differential evolution algorithm into the update process of harmony search algorithm by introducing linear decrement rule,and uses the tournament selection strategy to select the harmony in the mutation operator to improve the harmony search algorithm.Then,the improved initial harmonic weighting algorithm is used to select the optimal initial weight value of the wavelet neural network,and the wavelet neural network with fixed initial weight is trained to complete the load prediction model.The validity of the above model is verified by the actual load data.Secondly,the construction of short-term electric load forecasting model based on weighted EMD for wavelet neural network is studied.The power load sequence is decomposed into multiple sub-sequences by EMD decomposition;Clustering is performed according to the mean and variance of the subsequences,and the sequence with small mean and large variance is used as the standard sequence,and the reciprocal of the Euclidean distance between other sequences and the standard sequence in the category is calculated as the weight;Finally,sub-sequences are predicted separately by the improved wavelet neural network,the final prediction result is the weighted summation of each sub-sequence,and the model is verified based on the instance data set.Finally,the construction of short-term power load forecasting model based on blending model is studied.In view of the shortcomings of the instability of the prediction results of a single model,the combined model can flexibly utilize the advantages of different methods and can effectively improve the prediction performance.The XGBoost?RandomForest?Extreme Learning Machine and wavelet neural network are merged by blending framework,and the model is verified based on actual load data.