Comparison of Training Function, Adaption Learning Function, and Transfer Function of Hidden Layers in Artificial Neural Network in Weather Prediction

Weather prediction plays an important role in human life, such as agricultural commodities, aquaculture, trade, industry, tourism systems, airport/flight systems, mining industry, etc. This study is to compare the performance of several Training Function, Adaption Learning Function and Transfer Function in Artificial Neural Networks for weather prediction. The data used is historical weather that occurred in Semarang City, Indonesia from January 1, 2012, to December 31, 2022, as training data. Meanwhile, the data used as test data is historical weather data that occurred during the period from January 1, 2023, to December 31, 2023. The weather data consists of average temperature, humidity, and wind speed. Based on the training results, using trainlm, learngdm, and tansig which was trained using 30 hidden layers and 1000 epochs has the lowest MSE value of 5.08. Apart from that, the regression value of the training data also showed good results. That is, 0.99091 for details, 0.99102 for training, 0.98922 for validation, and 0.99263 for testing. The ANN architecture that has been trained using the configuration mentioned is used to make predictions on test data. The test results show that the ANN prediction error for temperature is 1.45% and humidity is 6.68%. Based on the results, it is proven that comparing three key factors of ANN for weather forecasting has a significant impact on the weather prediction ability of Artificial Neural Network.