Search

Data processing and integrated forecasting strategy has always been a major obstacle to the development of wind power forecasting system. In view of this, a novel decomposition method with SSA and CEEMDAN is constructed to decompose the original data, and also a sample entropy parsimonious integration model is applied to achieve ultra-short term wind speed prediction. Considering the respective data characteristics of each subsequence, we divide the decomposed multiple subsequences into three parts: high complexity group, low complexity group and residual group. PSO-ELM, IHOA-LSSVR, and IHOA-LSTM are applied to predict them respectively. Compared with other models with high accuracy in this paper, our model has higher prediction accuracy. [ABSTRACT FROM AUTHOR]

Wind energy, as new energy, becomes important support in the low-carbon transformation of the power industry. However, in wind farms, wind turbine fault diagnosis based on small-scale data is a thorny problem. To this end, this paper advances a wind turbine fault diagnosis method given feature selection and stacking model fusion. For unbalanced data, smote oversampling method is used to the effective instance of small class data and increase its proportion. RFECV is used to rank the importance of features, and then the features with high correlation are deleted according to the thermal map to reduce the dimension and obtain the feature subset. Then, XGBoost and LightGBM models based on 6-fold cross-validation is used to train the filtered data. To further improve the stability and generalization ability, a stacking fusion model based on logistic regression is trained using logistic regression. The results of this experiment are compared to other traditional methods by accuracy, ROC, and other indicators. The experimental results show that the strategy used in this paper has more satisfactory accuracy results than the traditional methods and can be used in wind turbine fault diagnosis engineering. [ABSTRACT FROM AUTHOR]

The chaotic nature of wind speed will damage power system seriously, and cause economic losses. Therefore, timely wind prediction is crucial for the safety of power system. However, the traditional prediction method is hard to fully learn the characteristic of wind speed. This paper proposes an optimal component IGSCV-SVR ensemble model to predict ultra-short-term wind speed. It changes the traditional single parameter optimization method of time series prediction. Firstly, the VMD based component correlation is applied to decomposing the original wind speed dataset to obtain multiple subsequences. Our model can find the dissimilarity of each subsequence, and then the model fully learns the feature of each subsequence. It can help improve the overall efficiency of ultra-short-term wind speed prediction accuracy. Finally, estimates are obtained by summing the prediction of all components. The case study proves the feasibility of our method through the comparative experiments with some previous prediction models in MSE, MAE, MAPE and running time in the experimental part of this paper. [ABSTRACT FROM AUTHOR]

In recent years, with computer and Internet technology developing at a breakneck speed, ensuring the fair dispatching and stable operation of the smart grid has increasingly become the focus of power companies. The emergence of the smart grid makes it difficult for a single model to accurately predict the complex power data, and the combined prediction model has become the main target of research by experts and scholars. To this end, a new combination prediction model is proposed in this research. The missing values in the data set are filled by the forecasts of the Random Forest (RF) approach, then the significance of input attributes is computed. After that, the feature selection is performed, and the peak load forecasting is modeled by using support vector regression based on particle swarm optimization (PSO-SVR). The experimental results using real data from a county in Jiangxi Province indicate that the model's prediction ability with feature selection is better than that of the model without feature selection. [ABSTRACT FROM AUTHOR]