Prediction of meteorological parameters using inverse artificial neural networks.

Abstract

Meteorological parameter data needed for climate change analysis, and monitoring the mitigation or adaptation measures taken, is often difficult to obtain due to the high costs of buying, installing and maintaining measurement equipment. This has resulted into data gaps. In this study, inverse artificial neural network (ANNi) algorithms were developed for the cheap and fast retrieval of some of the essential meteorological parameters basing on measured global horizontal solar radiation only. This study utilised ANNi algorithms because they require fewer inputs compared to feed-forward ANN, resulting in lower implementation costs. The parameters considered were solar radiation, sunshine hours, relative humidity, maximum, average, and minimum temperatures, and rainfall. For each algorithm, a feed-forward radial basis function neural network (RBFNN) was constructed, trained, and tested to predict global solar radiation based on the other selected meteorological parameter(s). After training, the optimal neural network architecture was saved, and used in the ANNi to aid the retrieval of the meteorological parameter(s). The inverse retrieval part of the algorithms employs non-linear optimization to retrieve meteorological parameters. We validated the retrieval algorithm using measured data that was not part of the training data used for the RBFNN, and several statistical metrics. We developed three meteorological parameter retrieval algorithms:- (i) A one−parameter ANNi algorithm which retrieves sunshine hours with correlation coefficient r, MnB, RMSE, and MAPE of 0.93, 0.056, 0.97 hr and 19.0%, respectively. (ii) The two−parameter algorithms most accurately retrieved the (SH, Tmax) pair with r, MnB, RMSE, and MAPE of (0.84, 0.87), (0.90, −0.001), (0.81 hr, 0.56 ◦C), and (12.1%, 1.6%), respectively, for each of the parameters. (iii) The ANNi algorithms for the simultaneous retrieval of three meteorological parameters most accurately retrieved the (SH, RH, Tav) set with r, MnB, RMSE, and MAPE of (0.73, 0.57, 0.61), (0.02, 0.13, 0.03), (0.97 hr, 13.69%, 1.61◦C), and (11.45%, 18.51%, 5.16%), respectively, for each of the parameters. The ANNi algorithms constructed in this research can improve weather forecasting and long-term climate monitoring in developing countries by predicting meteorological parameter values where only solar radiation measurement equipment is available. The data obtained can be used in several other applications such as in agriculture, civil aviation, and in the study of atmospheric energy balance, ecosystem evolution, and social sustainability