Explainable Spatio-Temporal Graph Neural Networks for District-Level Green House Gas Emission Forecasting in India

Abstract

Sub-national climate planning requires district-level greenhouse-gas (GHG) forecasts, yet emissions propagate across districts through electricity transfer, freight movement and industrial clustering that geographic adjacency does not capture. This paper develops ST-AGNN, a spatio-temporal graph neural network whose inter-district adjacency is learned as a convex combination of a geographic prior and a sparsifiable low-rank adaptive component, and evaluate it on a district-level benchmark (N = 120 districts, 240 months) whose latent spillover operator is known, allowing explanation quality to be scored rather than merely illustrated. Three findings emerge. First, adaptive structure learning improves accuracy: ST-AGNN attains MAE = 0.1351 Mt CO₂e month⁻¹ (R² = 0.9561) against 0.1538 for a static geographic graph, 0.1481 for a seasonal naive predictor and 0.3250 for a purely temporal GRU, with Diebold–Mariano statistics of 12.10 and 18.06 respectively. Second, and contrary to the premise motivating adaptive graphs, this predictive gain does not constitute structural identification: edge attributions recover the withheld non-geographic couplings with AUC = 0.456, indistinguishable from chance and below a simple lagged-correlation criterion (0.647). Third, counterfactual mitigation scenarios built on the fitted model are unreliable, because the renewable-share lever correlates +0.810 with emissions within districts through a shared time trend and the policy targets lie outside the support of the training distribution (only 1.6 % of observations exceed a 60 % share); the model consequently predicts that decarbonisation increases emissions. We conclude that predictive improvement from learned graphs should not be reported as evidence of recovered physical structure, and that correlational spatio-temporal models are inadequate for the counterfactual policy analysis they are increasingly used to justify.