This thesis is organized into three major stages, supported by two preliminary conference communications and four scientific articles, to address the sizing and management of hybrid renewable energy systems (HRES) from the perspective of complexity reduction, improvement of weather forecasting, and adaptive decision intelligence.
In the first communication, we analyzed the influence of economic parameters (capital, installation, and operating costs) and contextual factors (load profiles, local renewable-resource availability, feed-in tariffs) on microgrid design, revealing the essential trade-off between profitability and reliability. Building on this foundation, we proposed a hybrid method combining the Branch Bound algorithm with k-Nearest Neighbors to prune the configuration tree of PV+WT+BESS+DG systems within Simulink. From a synthetic meteorological dataset and load profiles, we generated 5,390 configurations—each characterized by vectors of load duration, predicted production, and penetration rate—and filtered them using Branch Bound to select an initial subset. A kNN algorithm then classified the remaining configurations by similarity, eliminating unpromising branches. This approach reduced simulation time by 45 %–95 % while maintaining over 83 % accuracy in identifying the optimal topology.
The second communication introduced a simplified prototype coupling a genetic algorithm with an LSTM model for global horizontal irradiance (GHI) forecasting, demonstrating the feasibility of combining statistical modeling with evolutionary computation. From this insight, we focused on GHI prediction via neural architecture search (NAS) enhanced by transfer learning (TL), dynamic search-space adaptation (DSS), and learning-curve extrapolation. Over one hundred candidate architectures—each defined by multiple hyperparameters (number of layers, kernel size, learning rate, etc.)—were explored. DSS dynamically refines the search space according to observed error distributions, enabling early abandonment of costly architectures, while learning-curve extrapolation halts networks with poor convergence. Compared to classical methods (GA, PSO, DE, ABC), this strategy reduced NAS runtime by up to 89 % and improved GHI-forecast accuracy by 33 %–99 % in RMSE over horizons of 6 to 72 hours.
The final stage introduces a comprehensive adaptive sizing and multi-objective control framework based on deep reinforcement learning (DRL). The state space incorporates battery-state-of-charge, forecasted irradiance and wind speed, and current generation levels, while the action space covers power dispatch among PV panels, wind turbines, battery, diesel generator, and grid. The Twin Delayed Deep Deterministic (TD3) algorithm ensures stability and efficient exploration. The cumulative reward combines the levelized cost of energy (LCOE), renewable-energy fraction (REF), and loss-of-power-supply probability (LPSP), with a penalty on fossil-fuel usage. Applied to real NREL load profiles, this approach reduced LCOE by 21 % to 30 %, increased renewable-energy share by 86 %, and decreased LPSP by 8.9 %, with peak performance of –19.7 % LCOE, +86 % REF, and –8.9 % LPSP—outperforming NSGA-II and MOPSO on most indicators. Together, these contributions illustrate the value of a progressive methodology spanning complexity reduction through simulation pruning and adaptive decision intelligence. The ultimate goal is to establish a design framework for next-generation smart microgrids that are resilient, high-performance, and environmentally friendly.
| Date | 30 Jan 2026 |
|---|
| Original language | French |
|---|
| Awarding Institution | - École de technologie supérieure
|
|---|
| Supervisor | Tony Wong (Supervisor) & Louis-A. Dessaint (Co-supervisor) |
|---|