This section highlights the specific research gaps that have inspired the dissertation in 5 sections:
1. Forecast-Optimization Integration Gap:
Current research treats forecasting and optimization as a decoupled sequence. Models used to forecast generate scenarios, which act as static inputs to optimization models. However, such an approach does not take into account the propagation of uncertainty and two-way feedback. These two features imply that insights from the optimization process can inform the forecasting one, which is generally not the case in practice. To our best knowledge, there is no operational framework through which iterative refinement between forecasting and optimization can be done for energy transition planning.
2. AI-Based System Gap for Turkey:
Turkey energy system modeling has mainly utilized deterministic optimization tools. These studies are useful but share common flaws. First, they rely on expert defined scenarios, rather than allowing the algorithm to search the solution space. Second, they treat uncertain parameters deterministically. Third, they do not possess the proper technology granularity. Fourth, they are not validated for long-term forecasting accuracy.
As far as we know, **no complete AI (artificial intelligence) based system** that combines neural-physical forecasting with multi-agent reinforcement learning exists for Turkey's energy transition. According to Evans [evans2016deepmind], DeepMind's work on optimizing the cooling of data-centres is an international application. Further, there is the instance of Google's work on forecasting wind power by using machine learning [sims2020google]. It should be noted that these activities certainly show what AI can do. But they are not at very large or long scales. In other words, they are for minutes to days and not for decades at national scale.
3. Long-Term Planning Gap:
Applications of RL to energy systems are mainly operational (minutes to days) and often geographically constrained. There is a notable lack of applications in the area of energy transition planning. When objectives are non-convex, the most common approaches and tools (MILP, stochastic programming) become intractable for state spaces of 446 dimensions, action spaces of 119 dimensions, and horizons of 26 years.
4. Explainability Gap:
AI systems are opaque, and they create key trust obstacles. Decision-makers seek clear rationales rather than simply optimal solutions. However, explainability in existing studies is lacking. There are few applications of SHAP analysis and counterfactual reasoning to optimize energy transitions.
5. Validation Gap:
Numerous studies fail to provide rigorous statistical tests, sensitivity analysis, and uncertainty quantification. It is very rare to see researches with proper temporal cross-validation. For example, in the k-fold cross-validation, the temporal ordering can be violated. We did not find any studies that comprehensively benchmark RL algorithms and forecasting models.
We address these gaps by the formulation of an integrated hybrid artificial intelligence framework that consists of (i) neural-physical forecasting models with physics-informed neural network (PINN) constraints and S-curve technology diffusion theory; (ii) reinforcement learning optimization system validated through comparison across multiple algorithms in various categories; (iii) bidirectional forecast-RL integration for uncertainty propagation and iterative optimization; (iv) rigorous statistical validation involving comprehensive testing and analysis; (v) transparency and reproducibility by way of open-source code, extensive documentation and detailed reporting of methodology. Details of our methodology like the number of algorithms, number of statistical tests, and technical implementation details are in Chapter 3 (Methodology).
Methods using machine learning, which combines physics-informed models with other models, strike a balance. Neural networks that apply physical laws like differential equations, conservation laws and constraints either in the loss function or the architecture are called Physics-informed neural networks (PINNs). Incorporating differential equations or physical constraints directly into neural network loss functions is why PINNs are so useful. Thus, models built with PINNs will follow fundamental physical laws while using being data driven through components of neural networks. Hybrid neural-physical models introduce physical constraints into neural forecasting architectures. Examples include energy balance equations and capacity factor bounds. Compared to purely data-driven models, applications to various domains have shown superior generalization. Nevertheless, there has been little work done on the adaptation of PINNs to energy capacity forecasting, even though these could evidently be used to enforce energy balance equations and capacity factor constraints. Chapter 3 (Methodology), provides the detailed discussion on PINN methodology and its application to energy forecasting.
Turkey’s energy system characteristics require country-specific AI models. Factors include vertically integrated state-owned enterprises, emerging competitive markets, diverse generators and resources, high import dependence with limited reserves, and financing constraints typical of developing economies. Generic international models miss these nuances, so context-specific development is necessary.
This research is important because Turkey has a short window to act. According to the International Energy Agency [iea2021netzero], the 2020-2030 decade is a determining decade for climate. It is which means that the infrastructure deployed in that decade will operate for 30-50 years, effectively locking in emission trajectories. Turkey's 11th Development Plan (2019-2023) and 12th Development Plan (2024-2028) will define investment directions and policies for the coming thirty years. As a result, evidence-based, scientific instruments used for planning are not just an academic exercise but rather a key instrument in policy affecting 85 million people and the global climate system.