递归神经网络在降雨量预测中的应用研究

Forecasting Rainfall with Recurrent Neural Network

递归神经网络(RNN)模型近年来在许多任务上表现出了优良的性能。运用具有长短期记忆(LSTM)单元的递归神经网络构建模型和通过时间反向传播(BPTT)算法更新网络权重解决长期降雨量的预测问题,较好地解决了高维数、非线性和局部极小问题。选取了前馈神经网络模型(FNN)、小波神经网络(WNN)模型和整合移动平均自回归(ARIMA)模型3种模型进行验证比较。仿真结果表明,递归神经网络模型优于其他模型,训练结果与实际值接近,预测精度较高。预测结果为农业用水管理、合理制定灌溉制度提供了重要的科学依据。

Recurrent neural network （RNN） model has recently demonstrated state-of-the-art performance in many tasks. This paper used recurrent neural network with long short term memory（LSTM） unit to model and back-propagation through time（BPTF） algorithm for updating network＇s weights to deal with the problem of long-term rainfall forecasting, which solved the problems of high dimension, nonlinear and local minima. Finally, three networks, feed-forward neural network（FNN） model, wavelet neural network（WNN） model and auto-regressive inte- grated moving average（ARIMA） model, are also provided for comparison. Simulation results demonstrate that recurrent neural network model outperforms the others networks, the training results are close to the actual value, the prediction accuracy is higher. The results provided im- portant scientific basis for agriculture water management, rational irrigation system decision.