Electric Load Forecasting Model And Bias-Variance Analysis

Reasonable planning of electric power system helps to lower costs and gains great social benefits.In order to maintain the effective operation of the power system,electric load forecasting is the key.In this thesis,the medium term electric power load forecasting and ultra-short term load forecasting are explored.In this thesis,the method of variance decomposition is studied.This thesis explores the nature of bias-variance in two aspects;classification prediction and numerical prediction.Two novel variance decomposition algorithms,one for regression problem and the other for the 0/1 problem,are proposed.The variance decomposition of the regression problem can decompose prediction error of the numerical prediction into the bias,variance and noise.The decomposition of 0/1 problem can divide classification results of the classification into the bias and variance.We investigate the influence of different algorithms,dataset size,training set size and running times on the prediction models.The experiment results show that:(l)Bias and variance decreased with the increase of dataset size;(2)Bias and variance can be used as a model evaluation indicator as the other traditional performance measurements;(3)When the training set and testing set ratio is 1:1,the accuracy of model is the worst;(4)When the size of dataset is bigger enough,reaching a certain point,the effect of training set size can be minimized;(5)When building prediction models,10 running times is sufficient to meet forecast requirements;(6)Different algorithms have different bias and variance.Therefore,according to the characteristics of bias variance,we can find the most suitable algorithm,data set size,training set size,number of cycles and other conditions to establish low bias variance model in the future.This thesis also proposes a new prediction model,named multi-factor-additive model(MFA).Multi-factor addition model will divide relationship of load and weather into 4 different trend components,which are temperature trend component,work day trend component,holiday trend component and other trend component.Each trend component is captured by a regression model,and then the trend component regression model is added to get the final forecast value.Through rigorous experiment,MFA demonstrates the following advantages:(1)MFA provides more accurate prediction models than other models'.On EUNITE electric dataset,MFA achieves a better prediction result than the one which won the No.1 prize in the competition in Europe.On North America electric dataset,MFA is also superior to the other algorithms commonly used in the industry.(2)The MFA model is the fastest model among all since it is linear time complexity.(3)The MFA model is easy to understand.It can be explained from various factors which affect electricity usage.The combination model is common used for ultra-short term load forecasting.This thesis proposes two kinds of combination models,one is based on ANN neural network and the other is weighted combination.Both of them are based on Grey-ARIMA models.One proposed combination model,named Grey-ARIMA ANN combined neural network model,consisting of two single forecast methods into the final model.G(1,1)and ARIMA model are used to predict the sample data at the same time.The output from the two single models is used as neural network models' inputs.And true value of the corresponding moments of expected output is used to train the neural network.After training,the ANN combines forecasting models and generates the final prediction.The prediction error of neural network combinational method is smaller than that of single model(G(1,1)and ARIMA model respectively),which greatly improves the accuracy of short term load forecasting.A weighted combination model based on Grey-ARIMA is also proposed.It uses the"minimum squared error" criterion to construct an objective function.It calculates weight coefficient vector of each individual prediction method.Finally,it combines each model according to their weights.Through the experiment,it shows that this weighted combination model has higher prediction accuracy,compared with single prediction model.The above mentioned experimental results,some of them verify the related reports in the field,some of them are novel.The major contributions of this thesis are listed as follows.1?In this thesis,the error decomposition of the regression problem and the error decomposition algorithm of 0/1 are proposed,which can help to evaluate the classification algorithm and the numerical prediction algorithn in the future.2?Bias-variance is verified as a performance measurement to forecasting model as well as the other classical performance indicators.When it is not convenient to use other evaluation indicators,it can be used to predict the model,which is helpful to find a better model with the characteristic of bias and variance.3?The proposed novel mid-tern electrical load forecasting model,MFA model,is more accurate and easy to interpret with faster running time.4?The proposed neural network combination model based on Grey-ARIMA combines the advantages of three kinds of different models.The prediction accuracy is higher than the single model.5?The proposed weighted combination model based on Grey-ARIMA combined models with the "minimal error square" criterion.Its prediction accuracy is higher than.that of the single model.