The Research On Short-term Gas Load Combination Forecasting Model Based On Residual Error And Similar Day Load Correction

With the rapid development of intelligent gas pipe network construction, the load forecasting is more and more important. The short-term load forecasting is a basic task of city gas system, the data are of fundamental significance for project planning,operation scheduling, network maintenance and engineering technical analysis.Therefore it is important to seek effective short-term load forecasting method in order to improve the prediction accuracy.At first, this paper deeply analyzed the gas load characteristics of Shanghai, the periodic trends and the seasonality of short-term load. We also studied the main influencing factors which provided the basis for the determination of the input vector of the forecasting model. Accurate and reliable historical load data were essential for exact prediction. Thus a series of pretreatment technology were used to identify and correct the abnormal data, which could ensure the overall trend and smoothness of load curve and improve the prediction accuracy. All of these made data preparation for the next simulation experiments.Then we expounded wavelet BP neural network and wavelet support vector machine which both were based on Morlet wavelet kernel function. They could improve the defects of single BP neural network and support vector machine.Meanwhile, adaptive inertia weight and cat chaotic mapping were introduced to enlarge the global and local search scope. We adopted improved particle swarm optimization algorithm to optimize their related parameters. These improvements were good for building the optimized model, speeding up the convergence rate and reducing the forecasting error. All of which made theoretical preparation for the establishment of the combination forecasting model.Finally, we built up the forecasting model and carried out mass simulation experiments. Because the gas load exhibited different characteristics in different seasons, seasonal alternation and holidays, we adopted three kinds of prediction model to cope with different situations in order to improve the overall predictive accuracy. According to the obvious seasonal characteristics of gas load, a residual error correction model was established to respectively forecast the load of four seasons. The specific method was: wavelet BP neural network made preliminary estimate, wavelet support vector machine made residual error correction, and the sumwas final result. As for the large deviation existed in seasonal alternating load forecasting, a weighted average model was built to forecast the load of alternate-season. The specific method was: the model’s outcome was the weighted sum of seasonal predictions. Because the fluctuation of holiday’s load curve was different from working days, a similar day load correction model was constructed to forecast holiday’s gas load. The specific method was: filtered history similar day within the comprehensive similarity, modified their load value based on the different relationship between them and the predict holiday, then assessed the load in terms of the comprehensive similarity and the adjusted load. Verified by a lot of experiments,the above models had better effectiveness and superiority for the gas load forecasting under different working conditions.