Dynamic optimisation of water allocation for hydropower and agricultural productivity in a Sri Lankan reservoir system under future climate variability

Abstract

At the global scale, increasing climate-driven hydrological variability is undermining the reliability of historical reservoir operating rules and intensifying competition between hydropower and irrigation, particularly in monsoon-dominated regions, as highlighted by the Inter-Governmental Panel on Climate Change (IPCC) Sixth Assessment Report. This study develops and applies a dynamic, machine learning–based framework to optimise water allocation between hydropower generation and agricultural irrigation in the Samanalawewa– Walawe reservoir system in Sri Lanka under current and near-future climate variability. The work addresses the growing conflict between energy and irrigation demands, compounded by increasingly variable monsoonal inflows, recurrent droughts, and the limitations of conventional rule-curve and programming-based reservoir operation methods in handling nonlinear, uncertain hydrological regimes. The methodology integrates three main components: (i) a daily rainfall–runoff model for the Samanalawewa sub-catchment developed in the Hydrologic Engineering Center–Hydrologic Modeling System (HEC-HMS); (ii) a machine learning module in which a Physics-Informed Neural Network (PINN) and a Long Short-Term Memory (LSTM) network are trained to simulate daily storage change based on the full reservoir water balance and are comparatively evaluated; and (iii) a Python-based dynamic water allocation model that embeds the best-performing ML architecture to generate daily irrigation and hydropower releases, subject to operational rule constraints. The HEC- HMS model reproduces historical inflows with Nash Sutcliffe Coefficient (NSE) values of 0.74 and 0.75, an acceptable Mean Ratio of Absolute Error (MRAE), and realistic flow- duration characteristics, indicating reliable representation of monsoon-driven runoff processes. Under normal hydrological conditions, both LSTM and PINN achieve high predictive skill for daily storage change (NSE ≈ 0.87–0.92), but under hybrid datasets including extreme events, the LSTM exhibits smoothed peaks and reduced accuracy (testing NSE ≈ 0.67), whereas the PINN maintains higher robustness (testing NSE ≈ 0.80) and superior mass-balance consistency, as shown by lower residual variance and more physically plausible parameter sensitivities. The dynamic allocation model, driven by observed inflows and climate-conditioned inputs, yields an optimised storage trajectory that remains consistently below the historically observed storage while satisfying minimum irrigation and hydropower thresholds, revealing a persistent surplus of about 30 MCM that can be reallocated to enhance irrigated area or energy production and reduce spill and evaporation losses. Future inflow and evaporation scenarios for 2021–2040 are generated by coupling downscaled CNRM-CM6-1 precipitation and temperature projections under two Shared Socioeconomic Pathways (SSPs), SSP2-4.5 and SSP5-8.5, with the calibrated HEC-HMS and an LSTM-based evaporation model, while irrigation and hydropower demands are projected using data-driven models conditioned on climate signals. Future projections (2021–2040) reveal that, while SSP2-4.5 maintains comparatively smooth inflow and release patterns, SSP5-8.5 produces sharper inflow peaks, longer low-inflow recessions, and more episodic hydropower and irrigation releases, indicating a clear shift from seasonally regulated storage dynamics to a highly variable, event-driven operational regime