Machine Learning Approaches for TemperatureForecasting in Sialkot, Pakistan: A Comparative Study with Traditional Methods

Abstract

This study presents a comparative analysis of traditional time series and machine learning (ML) models for short-term temperature forecasting in Sialkot, Pakistan. Using a decade-long dataset (2015–2025) that includes daily temperature, humidity, and atmospheric pressure, univariate and multivariate models were developed. Classical methods such as Simple Exponential Smoothing and Holt-Winters yielded high error rates (MAPE = 27%). In contrast, the Artificial Neural Network (ANN) model achieved the best precision (MAPE = 2. 97%, RMSE = 2.83, R2 = 0.960), closely followed by LSTM and Linear Regression. The results highlight the strength of ML for localized weather forecasting, with implications for agriculture, urban planning, and the reduction of disaster risk.