Significance and Contributions

The dissertation contributes substantially to theory, methodology, and practice in advancing scientific knowledge on using AI in energy systems while offering practical implications for Turkey's decarbonization strategy. The importance is multidimensional for multiple stakeholder communities: academic researchers in machine learning and energy systems; policymakers who design national energy strategies; industry practitioners who implement projects; and civil society organizations that advocate for climate action.

This dissertation makes the following key contributions:

1. We present a **hybrid neural-physical forecasting framework** among the earliest in **integrating technology diffusion theory** (S-curve dynamics) with physics-informed neural networks for long-term energy capacity forecasting, achieving R²=0.9799 (generation) and R²=0.9131 (capacity) with calibrated uncertainty quantification. At a 26-year horizon, we find that our best benchmark, the LSTM, achieves a model fit with an R²=0.71. Compared to this benchmark, we find our model **35% improvement**. Furthermore, we show that our model significantly outperforms an additional **17 alternative models** with statistical significance (p<0.001).
2. The project focuses on **hierarchical multi-agent reinforcement learning** in chained environment with hierarchical actions. The developed novel hierarchical action space of **119 dimensions** with **26 technologies** in **7 separate regions** further opens the opportunity to optimize the energy transition pathway of Turkey over a period of multi-decadal.
3. This is **one of the first operational frameworks** known to us that enables feedback between forecast and optimization. The two components converge in **4-7 iterations** with the forecast-policy gap being reduced from **8 GW to 1 GW** (an **89% reduction**). This integration leads to a **10-20% improvement in feasibility** (97.34% renewable vs. non-integrated 94.12%) and **3-7% improvement in quality** as measured by the reward optimization.
4. To the best of our knowledge, this dissertation provides among the **earliest systematic comparisons** of **13 RL algorithms** and **17 forecasting models** for energy transition optimization. It shows that the MAPPO is optimal to optimize energy transition (**4× more sample efficient** and **3× lower variance** than other alternatives). Moreover, it shows that the ENSM is the best forecasting model (all benchmarks are significantly inferior to the ENSM with p<0.001).
5. To the best of our knowledge, this is **among the earliest Turkey-specific optimizations** of the 2053 net-zero pathway that provides actionable policy recommendations: **220 GW capacity**, **97.34% renewables** (exceeding the 90% target by **7.34 percentage points**), **14.30 MtCO₂ emissions** (90.6% cut, exceeding the 87% target by **3.47 percentage points**), **$226.79B investment** (0.35% annual GDP), and regional and annual breakdowns.
6. A total of **78 pairwise Diebold-Mariano tests** plus other tests like ANOVA and bootstrap analysis through a comprehensive validation framework is incorporated in a temporal cross-validation. The above-mentioned techniques involve Diebold-Mariano and sensitivity tests to maintain the rigour of the methodology.
7. The entire code base, docker containers, and documentation will be publicly released upon acceptance to create full reproducibility and adaptability to other emerging economies.
8. **Integrating Explainable AI:** Counterfactual Reasoning and SHAP analysis transforming black-box optimization into a transparent decision-support system for policy design.

Theoretical Contributions

TC1: Hybrid Neural-Physical Framework for Technology Diffusion Forecasting

To the best of our knowledge, this dissertation provides a comprehensive framework that combines technology diffusion [rogers1962diffusion]; [@bass1969new], physics-informed neural networks [raissi2019physics], and Bayesian uncertainty quantification to forecast energy capacity over the foreseeable future. Earlier the applications of physics-informed neural networks was done in fluid dynamics, heat transfer, etc. that has PDEs (partial differential equations).

The study extends PINNs to socio-technical systems governed by diffusion dynamics and algebraic constraints (energy balance equations), with R²=0.9799 (generation) and R²=0.9131 (capacity), representing a 35% improvement over the best benchmark and outperforms 17 alternative models with statistical significance (p<0.001).

The theory gets its new perspective from 3 innovations: (1) multiphase sigmoid functions to depict the different adoption stages (experimentation, acceleration, saturation) instead of a single S-curve in the classical model, (2) attention mechanisms to identify the time-varying importance of different features, as different factors (policy, cost, resources) dominate across these stages, and (3) an enforcement of hard constraints via output transformation and penalty functions which guarantee physical plausibility ($0 \leq CF \leq 1$,$K \geq 0$,$G = K * CF * 8760/1000$).

This contribution expands our theoretical understanding of how one can incorporate theory (physical laws, diffusion theory) with data, thus resolving the long-standing conflict between theory-driven and data-driven modeling. The framework defines principled methods for hybrid modeling that can be applied outside energy forecasting to any domain exhibiting S-curve dynamics constrained by the physical laws (epidemiology, telecom, agricultural technology, etc.).

TC2: Hierarchical Action Space Design for Multi-Decade Sequential Decision-Making

The hierarchical action space of 119-dimensions over six decision levels, contributes something new to the theory of reinforcement learning of long-horizon planning problems. In the past, RL studies on energy systems used low-dimensional action spaces. For example, building control systems and microgrid dispatching used action spaces in the order of 1-10 dimensions. Others [cao2021reinforcement]; [@zhang2019deep] treated all actions in a flat high-dimensional space.

The hierarchical design considers the domain structure where category chosen (solar vs. wind vs. storage) occurs at a higher order of abstraction than investment allocation that conforms with human planning cognition. This design has three theoretical benefits. First, we can improve sample efficiency since the action space can be seen as decomposed. Second, we might get enhanced interpretability; i.e., instead of having an action vector that offers no insights into the reasoning of the AI, we can have a policy-level explanation like "prioritize solar in Mediterranean". Third, we can get transfer learning; i.e., the upper levels (which denote the category of the action) could be more easily transferable across countries while the lower levels (which dictate the specific site or installation) might be more country-specific.

this dissertation builds upon the hierarchical reinforcement learning literature (Sutton et al., 1999; Barto & Mahadevan, 2003) which has focused primarily on robotics and game AI. We target domains such as infrastructure planning. The RL action spaces can be explicitly adapted to incorporate engineering and economic decision structures, thereby generating a template for application to transportation planning, water systems, telecoms networks and other infrastructure sectors, which typically have long horizons and hierarchical decisions.

TC3: Bidirectional Forecast-Optimization Integration Theory

The bidirectional integration framework which allows the propagation of uncertainty from forecasting to optimization, as well as information feedback from optimization to forecasting, is a theoretical advancement over the sequential decoupled approaches espoused in the energy systems literature [prina2020integration]; [2018]; [Ringkjøb]; [et]; [al.]. The framework formalizes three integration mechanisms:

(1) **Uncertainty Propagation:** The forecast distributions$P(K_{2050} | data)$are generated using Monte Carlo dropout. In other words, they are stochastic environment transitions$P(s' | s, a)$. This lets RL agents learn robust policies under forecast uncertainty. To put it differently, it helps them avoid brittle policies that assume deterministic futures.

(2) **Consistency Enforcement:** Also, consistency is enforced through iterative refinements so that we have consistency in the energy balance,$|G_t - K_t * CF_t * 8.76| < \epsilon$. Further, model that have low credibility due to any discrepancy between the forecast and optimization interface shall be avoided.

(3) **Iterative Refinement:** The forecasts are conditioned on the trajectories that result from the RL process and then RL re-optimizes with the updated forecasts until convergence takes place ($\Delta K_{2050} < 1\%$). The two-way information flow finds Pareto improvements that cannot be attained through one-way coupling.

The theoretical contribution states the conditions under which the iterative refinement converges (Lipschitz continuity of forecast and optimisation functions) and bounds on number of iterations required (say about 5-7 for typical parameters). This work contributes to the literature on coupled modeling systems, whose convergence is usually confirmed empirically and lacks theoretical strongly indicates. The framework applies more broadly to coupled human-natural systems (e.g., climate-economy; land use-hydrology; fisheries-economics) requiring consistently integrated predictive and prescriptive models.

Methodological Contributions

MC1: Largest Systematic RL Algorithm Comparison for Energy Systems

The evaluation of 13 algorithms (policy gradient, evolutionary, fuzzy, multi-agent, meta-learning) is conducted across five categories in the largest RL algorithm benchmarking for energy transition planning. Most energy systems studies of the past adopted a single algorithm (MAPPO or DQN) without a comparison of results [zhang2021reinforcement], whereas others compared 2 to 3 algorithms only from a single algorithm category [cao2021reinforcement].

this dissertation provides empirical performance profiles along five dimensions: target achievement rates, sample efficiency (timesteps to 95% performance), convergence stability (coefficient of variation across runs), computational cost (training time, memory), and solution diversity (entropy of technology mixes). The results provide evidence-based algorithm selection guidance. For high-dimensional continuous action spaces, policy gradient methods (MAPPO, SAC, TD3) dominate and yield much higher success rates of 95%+ while evolutionary and fuzzy methods yield success rates of 60-75%.

The evaluation of the algorithms uses a novel methodology consisting of 100 independent runs for each algorithm in combination with sophisticated statistical tests consisting of 78 pairwise Diebold-Mariano tests and ANOVA. Furthermore a Bonferroni correction was used along with Cohen's d effect sizes and bootstrapped confidence intervals. These statistical tests set a new benchmark for automated reinforcement learning evaluations in energy applications. The findings highlight the requirement for statistical significance testing: several algorithms exhibit similar mean performance but significantly different variances which influence risk in practice.

MC2: Comprehensive 17-Model Forecasting Benchmark