Improving Air Temperature Prediction With Artificial Neural Networks

The mitigation of crop loss due to damaging freezes requires accurate air temperature prediction models. Previous work established that the Ward-style artificial neural network (ANN) is a suitable tool for developing such models. The current research focused on developing ANN models with reduced average prediction error by increasing the number of distinct observations used in training, adding additional input terms that describe the date of an observation, increasing the duration of prior weather data included in each observation, and reexamining the number of hidden nodes used in the network. Models were created to predict air temperature at hourly intervals from one to 12 hours ahead. Each ANN model, consisting of a network architecture and set of associated parameters, was evaluated by instantiating and training 30 networks and calculating the mean absolute error (MAE) of the resulting networks for some set of input patterns. The inclusion of seasonal input terms, up to 24 hours of prior weather information, and a larger number of processing nodes were some of the improvements that reduced average prediction error compared to previous research across all horizons. For example, the four-hour MAE of 1.40°C was 0.20°C, or 12.5%, less than the previous model. Prediction MAEs eight and 12 hours ahead improved by 0.17°C and 0.16°C, respectively, improvements of 7.4% and 5.9% over the existing model at these horizons. Networks instantiating the same model but with different initial random weights often led to different prediction errors. These results strongly suggest that ANN model developers should consider instantiating and training multiple networks with different initial weights to establish preferred model parameters.