基于季节指数调整的循环神经网络风速时间序列预测

提出一种基于季节指数调整的神经网络风速预测方法。针对历史风速之间的非线性关系,运用神经网络非线性拟合能力并结合季节性指数调整对风速时间序列进行预测。通过时序图法和增广Dickey-Fullerd检验法判断时间序列的平稳性,结果表明该序列为非平稳序列。这种不稳定性说明时间序列中可能包含趋势、季节性、循环和不规则成分的一种或多种,为此采用时间序列分解模型对时间序列进行季节指数调整。最后采用LSTM 和GRU神经网络预测风速,得到了较好的预测结果,且与未调整的数据预测结果及加法模型季节指数调整后的预测结果相比,基于乘法模型季节指数调整的2种神经网络预测结果有更高的风速预测精度。

RECURRENT NEURAL NETWORK PREDICTION OF WIND SPEED TIME SERIES BASED ON SEASONAL EXPONENTIAL ADJUSTMENT

A novel method of wind speed prediction based on neural network with seasonal exponential adjustment is proposed. Based on the nonlinear relationships among historical wind speeds and the strong nonlinear fitting ability of neural network, the neural network combined with seasonal exponential adjustment is adopted to predict the wind speed time series. First, time series graph and augmented Dickey-Fuller method are used to test the stability of time series. The results show that time series is unstable. The instability indicates that the time series contains seasonal, trending, cyclic and irregular components. In this paper, this time series decomposition model is used to adjust the seasonal index of time series. Finally, LSTM and GRU neural networks are used to predict wind speed data, and the ideal prediction results are obtained. Compared with the results of raw wind speed data and the seasonal exponential adjustment with the addition model, the results of two neural network methods based on the seasonal exponential adjustment with the multiplication model achieve much higher accuracy of wind speed prediction.