Existing non-power-generating dams represent substantial untapped renewable energy potential across developing regions, yet retrofitting decisions remain challenging owing to the need to balance economic viability against hydrological uncertainty while preserving primary dam objectives. This study presents a hybrid framework integrating Long Short-Term Memory (LSTM) neural networks with Stochastic Dynamic Programming (SDP) to optimise hydropower retrofitting under climate uncertainty. The LSTM forecasts monthly inflows from hydro-climatic predictors, while SDP derives an optimal retrofit/no-retrofit policy through empirically learned, non-stationary transition matrices, rather than the stationary transitions conventional approaches impose from historical frequencies. Test was performed on 37 years (1986–2023) of CHIRPS-derived monthly inflow data for Doma Dam, Nigeria, benchmarked against ARIMA, Classical Markov Chain, naïve rules, and perfect foresight bound. The hybrid model produced a mean NPV of $16.15M (99.6% success), statistically indistinguishable from Classical Markov, ARIMA, and Perfect Foresight. Its practical advantage lies in downside-risk management rather than mean-NPV outperformance, with higher CVaR₉₅ and lower coefficient of variation than Markov. Electricity tariff dominates NPV variance, followed by discount rate and capacity factor. The optimal policy proved robust to forecast perturbations up to ±20%. A Pettitt test located 1989 as the point of maximum divergence in inflows, though not significant; the LSTM still captured transition dynamics distinct from historical frequencies. Forecasting transferability, tested on another basin (Ojirami Dam), showed a substantial single-split accuracy drop, though a much smaller drop under walk-forward cross-validation. Overall, results support the retrofit's economic viability and the LSTM's forecasting superiority over ARIMA, without establishing decisive economic superiority.