Rehana: Another concrete example would be a model that predicts the probability of congestion in a transmission corridor for the next hour. The input consists of time series from several wind and solar plants, the load from different areas, and interconnector flows. A classical XAI method might simply tell us: “Wind farms A and B have the highest feature importance for this forecast.” A causal explanation allows us to take the additional step of simulating counterfactual scenarios such as: Rehana: Yes, the power systems domain has also shown a growing interest in hybrid models and recent advances in time series reasoning (TSR), physics-informed neural networks (PINNs), and reinforcement learning approaches that address this need. These emerging trends can offer more robust alternatives and help developers and stakeholders alike strengthen their decision making by identifying the root cause of a system’s output. “What if we reduced wind farm A by 5% without changing anything else?” In this case, we might see the congestion probability drop from 40% to 25%. Doing the same for wind farm B, a 5% curtailment might only lower the risk to 38%, because most of its power flows through a different corridor. This kind of explanation aligns much more directly with operational decision-making because it does not just tell dispatchers and planners which signals correlate with the problem. It tells them which levers are most effective. Bertram: Exactly. Correlation can suggest relationships, but it does not explain them. In dynamic, high-stakes environments like power systems, relying solely on correlation-based explanations can lead to spurious conclusions. This is where causal methods become essential. Instead of just asking “which variables are associated with overloads?”, we ask: › “If we intervene and curtail wind farm A by 10%, how does that change the probability of congestion in line 23?” › “If we raise day‑ahead prices in zone B, how much will that actually reduce evening peak load, after accounting for weather and other factors?” Techniques such as Causal Shapley values build on this “what if we intervene?” logic. They wrap your predictive model in a causal framework that distinguishes factors you can reasonably treat as causes (like setpoints, dispatched power, market prices) from those that are mostly background conditions (like weather). In addition to showing you which variables matter, they reveal how much actively changing those variables would alter the risk. RETURNING TO OUR ROOTS TO STRENGTHEN OUR WINGS: ENGINEERING AI TRUST FOR POWER SYSTEMS Bertram: In more complex dynamic interactions, especially those found in oscillatory or highly coupled systems, extended forms of Granger causality and modern approaches like NeuroKoopman-based causal discovery or crossmapping coherence can help identify directional influence between time series. However, the key message is that causal methods help us reason in terms of interventions and counterfactuals, which is much closer to how operators and planners think about control actions. 34
RkJQdWJsaXNoZXIy NTkxNzY=