Enhanced Weather Prediction with Feature Engineered, Time Series Cross Validated Ridge Regression Model

A wide range of applications pertaining to the weather, such as forecasting, prediction of extreme events and atmospheric analysis uses time series data to estimate the future values of weather elements. Time series models were extensively used earlier in the scenario which are now being largely substituted with deep learning models. Advanced deep neural architectures are found to deliver better performances than multi-variate time series models in handling high dimensional data with large number of temporal multi-variate predictors. Despite the efficacy, deep neural networks become inappropriate in low data regime weather applications due to the complexity, design issues and overfitting. The primary objective of this work is to produce a light-weight solution using classic machine learning models for enhanced weather prediction. The weather data of Kozhikode district, Kerala, India over a period of twenty-three years is taken for the proposed work with maximum temperature as target variable. Fourteen Regression based machine learning algorithms including ensemble models were used to model the data for choosing the best one. Detailed feature engineering helped the models to exploit the underlying patterns and seasonality. On comparison, feature engineered ridge regression model surpassed other models with the best predictive performance. Growing-window forward-validation ensured the temporal ordering during assessment. Additionally, the proposed ridge regression model was evaluated against time series models- baseline model and Seasonal Auto Regressive Integrated Moving Average (SARIMA), the proposed model excelled with a reduced Mean Absolute Error (MAE) 1.13 and Root Mean Square Error (RMSE) 1.54.