基于卫星数据和人工神经网络的逐时太阳辐射模拟研究

RESEARCH ON HOURLY SOLAR RADIATION SIMULATION BASED ON SATELLITE DATA AND ARTIFICIAL NEURAL NETWORK

关键词:  太阳辐射  气象卫星  人工神经网络  优化算法  损失函数

Key words:  solar radiation  meteorological satellite  artificial neural network  optimization algorithm  loss function

摘要:

目的 地面太阳辐射数据的获取受限于稀少的太阳辐射观测站点，气象卫星可以提供较高的时间和空间分辨率数据，为模拟太阳辐射提供了新的数据源；引入人工神经网络模型模拟太阳辐射可以获取更精准的数据指导农业生产。方法 文章利用气象卫星数据构建模拟太阳辐射的模型，在晴天和有云2种天气情况下，分别借助随机梯度下降算法（Stochastic Gradient Descent，SGD）、均方根传播算法（Root Mean Square Propagation，RMSprop）和自适应动量估计算法（Adaptive Momentum Estimation，Adam）优化的BP神经网络方法构建不同输入变量组合的晴空模型和云空模型，选取最优的算法和模型；再综合晴空模型和云空模型的结果，得到全天空条件下的太阳辐射模拟结果。结果 （1）BP神经网络的优化算法中，RMSprop和Adam算法比SGD算法表现更好，损失函数收敛更快且更稳定，误差更小；晴空模型采用RMSprop算法优化的BP神经网络模型，相关系数（Correlation Coefficient，R）为0.950，均方根误差（Root Mean Squared Error，RMSE）为84.27 W·m^-2，平均绝对误差（Mean Absolute Error，MAE）为58.99 W·m^-2；云空模型采用Adam算法优化的BP神经网络模型，R为0.884，RMSE为124.13 W·m^-2，MAE为88.08 W·m^-2。（2）利用卫星数据和地面气象数据共同作为输入变量构建的模型相比于不利用地面气象数据构建的模型效果更优，R增加0.015，RMSE减小11.93%，MAE减小21.77%。在云空模型的输入变量中增加反映云透光性的云光学厚度也可以减小模型的误差，R增加0.018，RMSE减小6.45%，MAE减小7.41%。（3）在输入变量相同的情况下，晴空模型模拟地面太阳辐射的精度比云空模型更高，R增加0.066，RMSE减小32.11%，MAE减小33.03%。晴空较有云的天气状况更加稳定，模拟效果更好。结论 通过不同优化算法的选择，以及利用卫星数据和地面气象数据共同构建模型，可以显著提升模型的精度。模型可以应用于解决部分地区太阳辐射值缺失的问题，对于需要利用太阳辐射数据进行研究的农业领域具有重要的意义。

Abstract:

Since the acquisition of Global Solar Radiation (GSR) data is limited by the scarce solar radiation observation stations, meteorological satellites can provide data with higher temporal and spatial resolution, which provides a new data source for simulating GSR. Also, using artificial neural network models to simulate GSR can yield more accurate data to guide agricultural production. Therefore, meteorological satellite data are used to construct models for simulating GSR. According to different weather conditions of clear and cloudy sky, the BP (Back Propagation) artificial neural network model optimized by SGD (Stochastic Gradient Descent), RMSprop (Root Mean Square Propagation), and Adam (Adaptive Momentum Estimation) algorithms were used to construct clear sky models and cloudy sky models with different combinations of input parameters. Then, to select the optimal algorithm and model. Finally, the results of the clear sky model and the cloudy sky model were combined to obtain the GSR simulation results under all-sky conditions. The results showed that: (1) Among the optimization algorithms of the BP neural network, the RMSprop, and Adam algorithms performed better than the SGD algorithm, which could make the loss function converge faster and more stable, and the error was smaller. The clear sky model adopted the RMSprop algorithm, the Correlation Coefficient (R) was 0.950, the Root Mean Squared Error (RMSE) was 84.27 W·m^-2, and the Mean Absolute Error (MAE) was 58.99 W·m^-2. While the cloudy sky model adopted the Adam algorithm, R was 0.884, RMSE was 124.13 W·m^-2, and MAE was 88.08 W·m^-2. (2) The models constructed using satellite data and meteorological data as input parameters were better than models constructed without meteorological data. R was increased by 0.015, RMSE was decreased by 11.93%, and MAE was decreased by 21.77%. At the same time, adding the cloud optical thickness, which reflected the transparency of the cloud, to the input parameters of the cloudy sky model, could also reduce the error. R was increased by 0.018, RMSE was decreased by 6.45%, and MAE was decreased by 7.41%. (3) Under the same input parameters, the accuracy of the clear sky model for estimating GSR was higher than that of the cloudy sky model. R was increased by 0.066, RMSE was decreased by 32.11%, and MAE was decreased by 33.03%. The reason was that clear-sky weather conditions were more stable than cloudy-sky weather conditions, which made the simulation much better. Therefore, by choosing different optimization algorithms, and utilizing satellite data and ground meteorological data to construct models, the precision of the models can be significantly enhanced. These models can be applied to solve the issue of missing GSR values in certain areas, which is of great importance to the agricultural field that requires solar radiation data for research.