This paper asks a useful question: if a data center can move work to another hour or location, or simply curtail it, which kind of flexibility is actually valuable to the grid? The answer depends almost entirely on the grid. In PJM, a realistic 30/50/20 mix of firm (demand must be served immediately), flexible (demand can be shifted in time and location), and interruptible (demand can be reduced near-instantly) load cuts modelled annual system cost by 6.3% in 2028 and 19.4% in 2038. The value is mainly spatial: move compute out of one location, Dominion, and avoid local capacity and policy obligations, including roughly 4.4 GW of gas build in 2028 and 8.9 GW of nuclear in 2038. In Korea, the value is temporal: move compute into solar hours. This makes another 0.5 GW of PV worth building in 2028 and later replaces about 1.2 GW of gas and 0.3 GW of batteries, although the total cost savings are only 0.3% and 1.2%. Data-centre flexibility is not one resource. It is worth whatever grid constraint it can actually relax.

A VPP with 4,000 EVs needs to tell the system operator how much energy and regulation it can deliver each hour. Computing this aggregate feasible region directly becomes impractical once thousands of battery, charger and availability constraints are coupled across vehicles and time. The paper learns a compact substitute. It dispatches the detailed fleet under historical prices, records the optimal schedules, then uses inverse optimization to fit one or two virtual batteries that reproduce those schedules. Using PJM prices from July 2022, the authors train on 20 days and test on 10. For decoupled EV operation (each EV’s regulation capability is treated independently), normalized dispatch error is 1.8% for an outer approximation and 1.6% with one learned battery. For coupled operation (power and regulation are allocated jointly across the fleet), the outer approximation gives 13.6%, one learned battery 13.4%, and two learned batteries 7.5%.

PV variability creates challenges for microgrid operation. Accurate PV forecasting and efficient scheduling are essential for optimizing system performance and reliability. This study proposes an LSTM-based PV power forecasting model integrated with multi-objective scheduling for a grid-connected PV-BESS. The framework quantifies how forecast accuracy affects PV self-consumption ratio, grid energy cost, grid injection, and battery utilization. Three scenarios are compared: perfect forecast, persistence model, and LSTM-based forecast. Results: The LSTM forecast reduces RMSE by 6% vs persistence, increases PV self-consumption from 78.1% to 84.5%, and reduces grid injections by 82%. The analysis also shows trade-offs: higher battery throughput from improved performance may accelerate aging.

The GB Capacity Market ensures security of supply by paying generation, storage, DSR and interconnection to be available when needed. CM auction agreements support investment in capacity. As large generation (including most nuclear) reaches end-of-life and renewables scale up, gas-fired plants are expected to transition to primarily providing security of supply. These assets would then rely increasingly on CM revenue. Satisfactory Performance Days (SPD) test whether capacity can deliver. Capacity Committed CMUs must demonstrate performance equal to or above their Capacity Obligation for at least one Settlement Period on three separate days (one between 1 Jan and 30 Apr). CMUs that fail have payments suspended and must demonstrate three additional SPDs between 1 May and 31 July or face termination. The government now proposes reforming the SPD process to test for enhanced availability and responsiveness.

From the UK government's report "The Grid We Need Now" on AI deployment in electricity networks: "Government should set a strategic intent to move to full probabilistic, risk-based grid operations by 2035 and planning by 2036" The current problem: GB grid planning and operations use deterministic N-1 (or N-2) security standards with no probability weighting. If a component can fail, the system is designed to handle it regardless of how unlikely that failure is or what it would actually cost. The result is routine overpayment for grid headroom and services to insure against low-probability events whose cost doesn't justify full insurance, plus connection offers and reinforcement investments based on models that don't reflect today's complexity. AI changes this. Forecasting and continuous modeling of current and future system states at operationally useful speeds is now feasible for the first time. Example: Open Climate Fix's multi-modal model combines NWP forecasts, real-time satellite imagery and live PV readings to halve solar generation forecast errors for NESO. Saves £30m/year in reserve costs by capturing spatiotemporal correlation. Spatially-aware wind forecasting using GNNs could predict constraint risk at key boundaries with enough lead time to preposition storage and flexibility instead of reacting after constraints bind. NESO's Volta programme is building toward an AI tool that quantitatively evaluates probabilities and risks and presents operators with different actions to consider. A major step forward, but still optimizing within deterministic security standards. Shifting the deterministic security standards to a probabilistic, risk-based approach requires deep sector-wide work. The risk-based methodology must be robust and demonstrably better at managing the risk currently approximated by deterministic standards.

The is_urgent classifier of Laya seems overly chill. I guess the churn_risk classifier was trained on a different question altogether. Or am I using this model wrong here?

In their new white paper, Volue argue that competitive advantage in BESS increasingly depends on algorithmic optimization and trading. Building optimizers myself, I am probably biased. But in my opinion both this white paper and the market trends below paint this as an exciting field to work in. Volue argue that: - BESS hardware has become increasingly commoditized: turnkey storage equipment prices fell approximately 81% between 2017 and 2025, excluding EPC and grid connection costs. - Alongside grid and market access, trading skill is an increasingly important source of repeatable advantage. - The hard part is earning a profit repeatedly while adapting to changes in market rules, revenue streams, and asset makeup. - Trading decisions are interconnected. A commitment now changes the opportunities available later. Degradation also means higher revenue today can reduce lifetime value. Volue identify three reasons behind this shift in importance from hardware to algorithm: - Greater granularity, including 15-minute day-ahead products, creates more intervals to forecast and optimize. Continuous intraday trading also demands timely reactions. - Shallow frequency-response markets saturate. Great Britain’s experience shows why wholesale trading, particularly continuous intraday, becomes increasingly important as the BESS fleet scales. - More BESS can also compress wholesale spreads. Adaptable algorithms help capture remaining opportunities, but cannot prevent market-wide cannibalization. Algorithms also need to accommodate co-located assets and mixed portfolios, where shared grid connections add constraints. Volue conclude that forecasting, optimization, and execution increasingly drive competitive advantage. I would add that several trends support the longer-term opportunity for BESS with some interesting figures from Modo Energy's Q1 2026 German outlook: - Renewable generation is projected to grow by 150%, from 280 TWh in 2026 to 695 TWh in 2040. More solar deepens the midday price trough, creating opportunities to shift energy into the evening. - Electricity demand is projected to grow by 70%, from 605 to 1,035 TWh, as transport, heating, and industry electrify. This can strengthen evening peaks, although flexible consumption can also help flatten them. - Higher gas and carbon allowance prices can widen spreads between renewable-rich hours and gas-dependent evening peaks. Taken together, Modo project a German market supporting 40 GW of BESS by 2040 (13x 2026 capacity). Despite that expansion, their forecast has two-hour battery revenues settling at around 115,000 EUR/MW/year by 2030. Against projected declines in CAPEX, Modo find BESS to continue operating above required investor returns. So overall, an exciting and promising space to work in!

With robust optimization we look for the highest-profit asset dispatch under the worst price realization allowed by our uncertainty set: max_{sell volume, buy volume} min_{prices ∈ uncertainty set} profit(dispatch, prices) Vidan et al. look into robust optimization both for a thermal generator and for a battery. Their assets are always price takers. This does not mean that prices are fixed before the auction. It means that prices are external: The assets assume that their own orders do not affect the clearing price. So the exercise here is to find the best-performing dispatch schedule or, for the generator, price-dependent offering curves under the worst price realization allowed by the uncertainty set. The authors then check how the resulting schedules or market orders would have performed under historically cleared prices. The issue with robust optimization, in particular for batteries, is that it does not care about the probability of different price realizations. In fact, their inner adversary (the inner problem) does not select one of the historical price trajectories. It can construct a damaging combination of hourly price deviations within the uncertainty set and the robustness budget Gamma. Assume the uncertainty set permits low prices during the morning and evening peaks and high prices during the PV peak. The robust model must protect against this inverted price shape even if it has negligible probability and does not resemble any historically observed day. A battery schedule shaped by such an uncertainty set can then charge and discharge in the wrong hours when normal market prices materialize - potentially buying high and selling low. Vidan et al. therefore conclude that pure robust optimization should be avoided for daily bidding when the objective is expected profit. They recommend a scenario-based approach and consider stochastic optimization more suitable: Here we assign probabilities to possible price trajectories and maximize probability-weighted expected profit instead of optimizing only against the worst admissible price realization. Robust optimization still has a place when protecting against the worst case is itself the objective - for example in investment decisions where one adverse realization could jeopardize the project.

There's a direct link between gas prices and wholesale electricity prices. The marginal cost of thermal generators (CCGT, CHP) is directly linked to how much they pay for their fuel. So when wholesale gas prices go up, the same generator will increase their electricity auction offer price. And when these gas-fired generators are marginal in the merit order stack the market clearing price goes up for everyone. Holding all else constant, how does a 1 Euro/MWh gas price increase affect the average daily electricity price? And how does this differ between the European bidding zones? The frontier economics report "The Fundamental Drivers Of Wholesale Electricity Prices in Europe" delivers some interesting figures. Gas price impact on electricity prices in Italy > Germany > France, while some of the Nordics are inversely correlated. The latter is interesting and I should dig into how hydro probably plays a role here. What is also interesting is that there are in fact differences in the impact between neighbouring bidding zones, which likely points to congestion.