Multi-model forecasting framework for agricultural nutrient dynamics in India: a comparative analysis of ML and hybrid approaches for NPK consumption

Abstract

This study developed and evaluated a comprehensive forecasting framework for predicting the dynamics of agricultural nutrients (N, P2O5, and K2O) in India across three dimensions: consumption, exports, and imports. We implemented a diverse set of nine forecasting models, spanning traditional time series methods (ARIMA), machine learning algorithms (Random Forest, SVM, XGBoost), deep learning approaches (ANN, LSTM, GRU), and hybrid architectures (ARIMA–LSTM, XGBoost–LSTM. These were compared using historical data, and performance was analyzed with MAE (mean absolute error), MSE (mean squared error), and RMSE (root mean squared error). ARIMA performed consistently well in predicting trade in N and K2O, while advanced machine learning models like XGBoost and Random Forest excelled in forecasting agricultural consumption. Six-year-ahead predictions (2024–2029) indicate rising nitrogen consumption (65,027 tons to 69,845 tons), stable phosphorus usage (29,006 tons to 30,211 tons), and increasing potassium demand (20,807 tons to 24,301 tons). Our results suggest model-specific advantages for different prediction scenarios, with hybrid models providing negligible improvements over simpler approaches. This research offers valuable insights for agricultural planning, policymaking, and food security in India. The data used were obtained from authoritative sources, including the Food and Agriculture Organization (FAO) and the Fertilizer Association of India (FAI), ensuring reliability and national relevance.