Power Load Forecasting Analysis And Research Based On Machine Learning

The grid system is the energy source for urban operation and development.In order to effectively control the energy flow of the grid system and realize the dataization and intelligent operation of the grid system,short-term power load forecasting has become an important research topic to ensure grid system scheduling and status detection.Based on the problem of poor fitting ability of traditional mathematical statistics methods,this paper uses a variety of machine learning methods and actual data to predict the short-term power load to improve the model's nonlinear fitting ability.First,the preprocessing of load data and the construction of feature sets are studied.In order to standardize the input of the model,the missing values ??and outliers were first identified,filled and corrected.The accuracy of missing value filling is improved by using a combination of horizontal time-series filling and vertical similar time point filling,and the accuracy of outlier identification is improved by dynamically adjusting the confidence bandwidth.Secondly,through the historical load data and weather and seasonal information,the key factors influencing the load are fully mined,and then feature coding,removal of small variance features,removal of highly relevant features,tree model screening,etc.are used to standardize and extract features for the model Screen out the best features.Then,a dual neural network model based on an improved harmony search algorithm is studied.Aiming at the problem of slow parameter search speed and difficult convergence of standard harmonic search algorithm,an improved scheme of assigning weights to harmonics and adaptively fine-tuning the wideband is proposed.The effectiveness of the improved algorithm is verified by multiple test functions.Then the convolutional neural network is combined with the recurrent neural network,and the parameters are updated based on the improved harmony search algorithm.The example verifies that the dual neural network model based on the improved harmony search algorithm has a higher fitting ability,far exceeding Power station actual demand.Finally,the short-term power load forecasting method based on the stacking model fusion framework is studied.Aiming at the shortcomings of unstable prediction results of a single model,a dual neural network model,a random forest model,an XGBoost model,and a Light GBM model are used as the base model,and a linearmodel is used as the high-level model.A load forecasting method based on the stacking model fusion framework is established.Through comparison of experimental results,the fusion model has smaller prediction error and higher prediction accuracy than the single model.The research results of this paper have certain scientific significance and use value for improving the accuracy and reliability of short-term power load forecasting.