| Neural network has been successfully applied in many fields such as information processing, pattern recognition, intelligent control and system modeling since it was first proposed. Moreover, researches have used combination or hybrid model based on neural network algorithms for forecasting and got good results. This paper analyses effectiveness and robustness of neural network combination models in practical prediction by simulating electricity load and wind speed data.In a competitive electricity market, electricity load series has proved to be highly complex due to its nonstationarity as well as various unstable factors. Thus, how to address forecasting accuracy is an important challenge in an era in which electricity market is increasing significant. Based on Support Vector Machine (SVM), Adaptive Neuro-Fuzzy Inference System (ANFIS), Extreme Learning Machine (ELM), method of feature extraction and Optimization Algorithm (OA), this paper proposes Time Series Combination Optimization Model (TSCOM) for day ahead electricity load forecasting. This novel model successfully combines advantages of each basic model meanwhile considering characteristics of time series rather than simply giving each model a specific weight. This method of feature extraction makes OA can obtain a relatively decent weights to combine basic models and then improve forecasting accuracy. For the purpose of analyzing and validating forecasting effectiveness of TSCOM, the electricity load in Queensland, Australia is selected as database in this paper. Finally, the simulation result shows the proposed model indeed outperforms SVM, ANFIS and ELM.Considering to the wind energy application, wind power is difficult to implement at a large scale because the volatility of wind hiders the prediction of steady and accurate wind power or speed values, especially for multi-step-ahead and long horizon cases. Multi-step-ahead prediction is challenging and has rarely been studied by statistical and machine-learning methods because greater numbers of forecast steps correspond with lower accuracy. Although multi-step forecasting can be realized by the Weather Research and Forecasting Model (WRF), a large error in wind speed will result from inaccurate predictions at the beginning of the synoptic process. Due to the limitations and uncertainties of the multi-step wind speed forecasting method, this paper proposes a novel hybrid forecasting model, the application of the adjusted Lazy Learning model based on a denoising, decomposed multi-output strategy and improving the Validation Cuckoo search (ALL-DDVC). The proposed model is not only capable of multi-step ahead wind speed forecasting with abnormal data, but is also effective and robust according to valid experimental simulations of ten-min-interval wind speed data from four wind farms. Therefore, the ALL-DDVC method is practical and effective for multi-step and long-horizon wind speed forecasting. |
