Daily Minimum and Maximum Temperature Forecasting in Sri Lanka: An Artificial Neural Network Approach

Abstract

The National Meteorological Center, Department of Meteorology, Sri Lanka is not currently using technologically advanced methods in forecasting daily minimum and maximum temperature of selected locations in the country. In the city weather forecast, they mainly focus on ten cities namely, Anuradhapura, Badulla, Batticaloa, Colombo, Galle, Hambantota, Jaffna, Kandy, Ratnapura, and Trincomalee, covering the entire island. Motivated by the requirement for a sophisticated forecasting technique, we introduce an Artificial Neural Network (ANN) approach for this problem using previous weather data as inputs from more than ten locations in Sri Lanka over ten years (2010-2019). The data used in this work were obtained from the Department of Meteorology, Sri Lanka. A three-layer (input, hidden and output) ANN having appropriate number of nodes in each layer and with the Ward architecture was constructed which uses three activation functions (Gaussian, Gaussian complement, and hyperbolic tangent) in the hidden layer. The model was validated using the k-fold cross-validation procedure. The results, that is, daily minimum and maximum temperature, were obtained using the R software package (4.0.3 version). It was observed that the predicted values were very homogeneous compared to the real values with a small error and this error was reduced using the gradient descent method. We further investigated how various choices of the number of hidden neurons and the epochs affect these results. It was found that the best number of neurons in the hidden layer was twenty one and if the number of epochs was increased the error was approaching zero. A close agreement between the real and predicted temperature values were observed in this work.