Climate Change Prediction Utilizing Machine Learning Techniques

The most significant renewable natural resource is water. Water management is crucial for the sustainability of human life. One of the most crucial aspects of an area’s water management is the rainfall prediction. A collection of observations of a variable made at regular periods is referred to as a time series. On the other hand, a prediction is only a computation of what will happen in the future of the variable of interest using information from the past under the presumption that the pattern followed in the past would likewise be continued in the future. Using traditional and computational models, this effort will attempt to develop forecasting models for the time series dataset. This research will make use of yearly climate data for Varanasi city for a total of 112 years (1910-2022). Moving average, exponential smoothing with one parameter and the traditional model auto-regressive integrated moving average (ARIMA) are the distinct statistical models to be taken into consideration. Varanasi rainfall data is steady; consequently, there is no need to differentiate the data series. Nevertheless, the provided data was non-normal; as a result, the data was transformed to make it normal. This converted data was taken into account for all models, including hybrid models, when further analysis was performed. When using moving average, forecasts was made for several time periods and the best one was selected using error measurements. Individual forecasts were also performed using the computational model k-nearest neighbour (KNN) and the interpolation technique cubic spline. In the instance of a cubic spline, many knots were tested, the best value was picked and analysis was performed based on the chosen knot. Finally, the best statistical models and the interpolation model spline were combined with kNN to create hybrid models and these hybrid models were used to forecast the provided data for the years 2018-2023. All models, both individual and hybrid, were compared and the best model from each was chosen based on the error metrics mean absolute percentage error (MAPE), mean absolute deviation or mean absolute error (MAD/MAE) and root mean square error (RMSE).