Posters

Abstract

Ultra-short-term wind power forecasting is very important for power systems in the context of high wind power penetration and power market transactions. In recent years, a large number of prediction methods based on machine learning have been applied to the field of short-term wind power prediction, which has greatly improved the accuracy of prediction. In this research, a hybrid intelligent ultra-short term wind power forecasting algorithm is proposed, including three parts: the data pre-processing, the hybrid training stage, and the module's optimization. The first step in the pre-processing data stage is data cleansing to eliminate the missing data points and outliers in the original wind power sequential data. Additionally, wavelet transform is introduced to separate signals in the frequency domain to increase the stationarity and predictability of the time series. Then, the time series data set will be standardized and distributed into the training set and testing test for training, validating, and testing the models. In the hybrid training stage, three-time series prediction models are introduced. The Recurrent Neural Network(RNN), Long Short-Term Memory(LSTM), and mixed input features-based cascade-connected artificial neural network(MIF-CANN). The three models are trained independently and give prediction results separately. By comparing the prediction results generated by the models, the compatibility between the model and the data set is revealed. A final decision has not been made at this stage, although a vague conclusion can be reached. In the optimization stage, parameter tuning is in the process of improving the performance of each model. The Adaptive Network-based Fuzzy Inference System (ANFIS) is introduced to learn the mapping between the input and output of each model. The ANFIS is instructed to avoid the errors caused by the instability of model performance because different models trained by the same data set still vary on performance because of their ability to extract various features. The self-adaptive and self-learning ability of the algorithm can modify parameters in the models based on the changes in prediction results. In the meantime, the Genetic Algorithm (GA) is introduced to optimize the performance of the ANFIS by tuning its parameters to avoid local optimum. An adjusted prediction result is generated in the end. In conclusion, the hybrid intelligent forecasting algorithm has decreased the standardized root-mean-square deviation and standardized mean-absolute error. It can significantly improve the prediction accuracy and solve the problem of the significant prediction error of high-frequency eigenmode function in empirical mode decomposition. Compared to a single model method, the prediction error has decreased by 2%. In the meantime, by decreasing the max prediction error in each time series prediction model, the adjusting error methods in the optimization stage have increased the algorithm's robustness. The ANFIS using GA optimizing has provided the algorithm with self-adaptive error correction. It can combine and evaluate the prediction results from all models, and the average MAE has decreased 10%.