Machine Learning Models Track Rising Heatwave Magnitude in Western Maharashtra

Heatwave magnitudes in Western Maharashtra, India, have shown a consistent upward trend over seven decades, with the steepest increases recorded in the Kolhapur and Pune districts, according to a new study published in Theoretical and Applied Climatology. The research team, led by Nobin Chandra Paul, developed an integrated framework that both characterises historical heatwave behaviour and forecasts future magnitude using the Heatwave Magnitude Index daily, or HWMId.

The researchers constructed annual HWMId series for six districts spanning 1955 to 2025, drawing on gridded temperature data from the India Meteorological Department. Trend analysis across the districts revealed a persistent increase in heatwave magnitude, a pattern the authors link to growing risks for this climate-sensitive agricultural region.

A key element of the study was the detection of structural breaks in the time series. Using the Pettit and Buishand change-point tests, the team identified significant shifts in heatwave behaviour during the late 1990s to early 2000s, suggesting that the underlying statistical process generating extreme heat changed around that period. Before modelling, diagnostic checks confirmed significant autocorrelation via the Box-Pierce test and nonlinear dependence via the BDS test, which justified combining linear and nonlinear modelling approaches rather than relying on a single method.

Within this framework, the authors built four forecasting models: a standard ARIMA model, an Interrupted ARIMA model, an Artificial Neural Network (ANN), and an Interrupted ANN. The interrupted variants incorporate pulse intervention variables derived from the change-point analysis, allowing the models to account for abrupt shifts in the series rather than treating the whole record as statistically uniform. The team also applied hierarchical clustering based on innovative trend analysis indicators to group districts with similar heatwave characteristics.

Model performance turned out to be district-specific. The Interrupted ANN model delivered improved prediction accuracy for Kolhapur, while ARIMA-based models performed better for Pune. The Diebold-Mariano test, a standard procedure for comparing forecast accuracy, supported these district-level differences, with stronger statistical evidence of improvement observed for Kolhapur. This outcome underscores that no single model dominates everywhere, and that structural break information can sharpen nonlinear forecasts in some locations more than others.

The authors argue that integrating change-point detection with machine learning offers a more faithful representation of complex heatwave dynamics than conventional approaches. Because the framework is modular, it could in principle be extended to other regions where long temperature records and gridded data are available.

Beyond the technical contribution, the study points to practical applications in climate risk management, agricultural planning, and the design of early warning systems. The researchers caution that the framework’s value lies in its scalability: the same pipeline of trend analysis, change-point testing, diagnostic screening, and comparative model evaluation can be replicated for other vulnerable districts.

Subject of Research: Heatwave magnitude characterisation and forecasting using the HWMId index and hybrid time series and machine learning models in Western Maharashtra, India.

Article Title: Characterisation and forecasting of heatwaves using HWMId: Change-point based time series intervention modelling with machine learning for Western Maharashtra, India

Article References: Paul, N. C., Chanda, B., Ponnaganti, N., Rathod, S., Nangare, D. D., & Reddy, K. S. (2026). Characterisation and forecasting of heatwaves using HWMId: Change-point based time series intervention modelling with machine learning for Western Maharashtra, India. Theoretical and Applied Climatology, 157(11), Article 761. https://doi.org/10.1007/s00704-026-06698-7

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06698-7

Keywords: heatwaves, HWMId, Western Maharashtra, change-point analysis, ARIMA, artificial neural networks, time series forecasting, climate risk, India Meteorological Department, early warning systems, Characterisation, forecasting