The Load Forecasting Model Which Parameters Optimized By MCMC Based On Bayesian Theory

Electric load forecasting is one important component of electricity technical economy. It is a very important work for electricity network programming, electricity power system satisfyingly running and load distribution. This paper presents a new type of load forecasting model. Based on the Bayesian theory, the model parameters were optimized by the Markov chain Monte Carlo algorithms. During the learning process, given the sample data and prior distribution, The posterior distribution was used to estimate the model’s parameters. The posterior calculation is an integration over the high dimensional parameter space. Usually numerical integration algorithms were used to compute the posterior. In this paper, the high dimensional integration was approximate calculated by two MCMC algorithms Gibbs sam-pling algorithm and Metropolis-Hasting sampling algorithm respectively. The main contents discussed in this thesis are:1)The relationship between the weather factors and electricity load was discussed. An-alyzing the sample data, the scatter figures and fitting curves were used to determine the relationship. The conclusion is that temperature and relative humidity are two important weather influence factors.2) The load curve forecasting model is established based on Bayesian Neural network learned by hybrid Monte Carlo Markov chain algorithm. The weather factors and time vari-able are the input variables, and electricity load is the output variable. The hidden neural units were given artificially. In Bayesian Neural network, the weight vector parameters were learned by a new hybrid Monte Carlo Markov chain algorithm. Bayesian neural network is considered as a Hamilton dynamic system, the weight vector parameters are the position variable in the Hamilton dynamic system. The Leapforg iterative algorithm and Metropolis-Hasting algorithm combines to construct an Markov chain of weight vector parameters, such that the in variant distribution of Markov chain is the desired posterior distribution. Then, using the Markov chain, we can get the estimation value of the weight vector parameter. The hourly loads ahead seven days were forecasted by the Bayesian neural network. The experi-ment result shows that the Bayesian neural network learned by hybrid Monte Carlo Markov chain algorithm has higher forecasting accuracy and good generalization performance than artificial neural network learned by BP algorithm. It can welly overcome the over-fitting phenomena.3) A monthly special load forecasting method is proposed. The forecasting model is the state space model with explanatory variables. Because there exists some positive correlation between special load and weather factors temperature and relative humidity, the cooling degree and heating degree and relative humidity are the three explanatory variables. Our state space model has two types parameters, the regress coefficients of the explanatory variables and the variance of the fluctuation term. The regress coefficients are estimated by the Kalman filter. Gibbs sampling and Metropolis-Hasting sampling algorithms are used to construct a Markov chain of Variance parameters. Using Monte Carlo approximation calculating method, the conditional posterior expectation is obtained from the Markov chain. The estimation of variance parameters are conditional posterior expectation. The state space model can welly smooth the sample data. The root mean square error(RMSE) and Mean absolute percentage error(MAPE) are small. The State space model with weather factor explanatory variables forecasts the maximum load and minimum load ahead6months. The forecasting result shows that this model has higher forecasting accurity for the first three months, but for the second three months, the forecasting error is relatively large.4)This paper presents a new ultra short term load forecasting model-Bayesian ARIMA-GARCH model. The fluctuation of ultra short load data has the heteroscedasticity property. The ARIMA(p,d,q)-GARCH(1,1) model can simultaneously consider the change ruler of the mean item and the heteroscedasticity of the variance item. ARIMA(p,d,q) model describes the attribution of mean item and GARCH(1,1) describes the variance item. It’s parameters p, d, q are determined by the autocorrelation function(ACF) and partial autocorrelation func-tion(PACF). The coefficients parameters of GARCG(1,1)model usually are estimated by the pseudo-maximum likelihood. This paper propose a new combinational sample estimation method. Given the prior distribution and the sample observation data, An candidate sam-ple of variance parameters and free degree parameters is drew by the Gibbs algorithm form the conditional probability distribution, and the Metropolis criterion determines that if reject this candidate sample. Thus, the combinational sampling method can get a Markov chain of these parameters. After the parameter estimation work, the ARIMA(p,d,q)-GARCH(1,1) model which parameters are estimated by MCMC algorithm and PMLE algorithm respec-tively forecasts the twelve load value at every5minture time point of next hour. From experiment result, the ARIMA (p,d,q)-GARCH(1,1) Estimated by MCMC algorithm have higher forecasting accuracy than that estimated by QMLE algorithm. The forecasting errors RMSE and MAPE are far smaller than that of the single ARIMA model.