DEVELOPMENT OF HYBRID MODELLING FRAMEWORK FOR FLOOD AND DROUGHT PATTERNS UNDER CLIMATE CHANGE VARIABILITIES IN RWANDA

Abstract

Rwanda’s heavy reliance on rain-fed agriculture and its complex, steep topography renders it highly vulnerable to escalating climate-induced floods and droughts. Despite advancements in modern predictive technologies, the "black box" nature of traditional machine learning algorithms has significantly hindered their integration into practical policy making and water resource management. To address this gap, this dissertation develops and validates a novel hybrid modelling framework to analyze historical hydro-climatic trends (1981–2024), project future extreme scenarios up to the year 2100, evaluate socio-economic vulnerabilities, and formulate evidencebased adaptation strategies. The methodology leverages high-resolution Earth Observation Data (CHIRPS and ERA5-Land) alongside a bias-corrected CMIP6 Multi-Model Ensemble to force a SWAT-LSTM hybrid model. Within this framework, the physical SWAT+ model calculates the baseline water mass balance, while a Long Short-Term Memory (LSTM) deep learning network serves as a residual corrector to accurately capture non-linear routing processes and sub-daily temporal dependencies. Historical trend analysis reveals a severe "Temperature Penalty" across Rwanda, where the frequency of warm nights has doubled, dramatically elevating potential evapotranspiration (PET) and driving agricultural drought even in areas where annual rainfall volumes have recovered. Concurrently, the return period for 50-year extreme storm events in the Northern Province has contracted to just 5 years, frequently overwhelming existing municipal infrastructure. Through the integration of the LSTM network, the hybrid predictive model drastically outperformed standalone physical models, reducing the Root Mean Square Error (RMSE) by up to 74% and significantly improving the Kling-Gupta Efficiency (KGE) for discharge predictions. Future projections under the extreme SSP5-8.5 emission scenario indicate a critical "Precipitation Paradox". The model predicts a 15.9% acceleration in extreme flash flood peaks (reaching 250.80 m³/s) at the Ruliba catchment, severely threatening Kigali's urban infrastructure. In stark contrast, the downstream Rusumo catchment faces a +3.29°C temperature rise, generating an extreme evaporative demand of 1,460 mm annually that threatens the baseflow reliability of the Regional Rusumo Falls Hydroelectric Project and regional food security. To mitigate these geographically polarized hazards which are characterized by excess flood energy in the West and severe volume deficits in the East, the study proposes a "Spatial Rebalancing Philosophy". A Multi-Criteria Analysis (MCA) prioritized the construction of high-capacity Rainwater Harvesting (RWH) Catchment Dams, the deployment of LSTM-enhanced Dynamic Early Warning Systems, and the exploration of Induced Inter-Basin Water Transfers as the most effective adaptation measures. The successful execution of these strategies provides a data-driven blueprint to enhance Rwanda's climate resilience, actively supporting the National Strategy for Transformation (NST2) and directly aligning with the United Nations Sustainable Development Goals (SDGs 2, 6, 11, and 13).