When every quarter-hour counts: What Switzerland can learn from the UK

The Swiss electricity industry is moving into a new reality. With the massive build-out of photovoltaics, rising intraday volatility and Swissgrid's single-price imbalance mechanism in force since 1 January 2026, forecast quality is becoming a direct economic factor. What used to be a technical side process is increasingly becoming a lever for trading margins, balancing-energy costs and operational stability.

Britain has already gone through this transition. There the «Market-wide Half-Hourly Settlement» (MHHS) reform forces suppliers to predict precisely, for every half hour, how much electricity their customers actually consume. The same principle applies in Switzerland, here on the finer quarter-hour grid. Errors become financially visible immediately.

Over the past eighteen months CentraLogic in Britain — with whom NDG has jointly evaluated transfer to the Swiss market — has built an AI-driven forecasting infrastructure: in production, regulatorily relevant and tested under real market conditions.

The central question is therefore no longer whether AI forecasting models will become relevant in the electricity industry. It is how fast the Swiss industry can build the necessary capabilities.

Why forecasting becomes a Profit-&-Loss lever

Four structural developments are hitting the Swiss energy industry at the same time:

• The solarisation of the grid is changing load profiles and increasing residual-load volatility.
• Hydropower is becoming a strategic flexibility reserve in the European market.
• Coupling with EPEX and neighbouring countries is increasing intraday dynamics.
• Since 2026 Swissgrid charges balancing-energy deviations directly to balancing groups via a single-price mechanism.

This makes forecast accuracy measurable on the P&L.

The challenge does not only affect large corporates. Around 600 distribution network operators in Switzerland (predominantly integrated municipal and communal utilities that also supply energy) face similar problems, often with limited staff resources and heterogeneous data landscapes. The Digital@Utility 2026 study by Kearney et al. shows that around 70 percent of AI initiatives in the energy industry remain stuck at pilot stage and rarely move into productive operation. Many companies have formulated AI strategies, but productive machine-learning systems with clear KPIs remain the exception. The biggest hurdles are not missing ideas, but missing data infrastructure, talent and operational capacity.

What was actually built in Britain

The British platform is not based on a single «super-model» but on several specialised components.

Load forecasting works with separate models per customer segment: households, SMEs, heat pumps, electric mobility or industrial consumers. The reason is pragmatic: a single national model looked good on the averages but systematically mispredicted precisely the critical segments.

Price forecasting in turn deliberately separates day-ahead, intraday and balancing-energy models. Attempts with shared neural networks appeared more efficient but worsened predictions exactly where price errors were most expensive.

Price forecasting today works as a hybrid model: roughly 70 percent machine learning (ML) and 30 percent fundamentals-based sensitivity models for gas, CO₂, generation mix and cross-border flows. The combination of the two model approaches is continuously corrected for systematic biases. The reason is sobering and simple: pure ML models win the average week but often lose during market phases with structural breaks. Yet those very phases are the most expensive.

Quantile regressions were also introduced to produce not just point forecasts but robust confidence bands. Particularly during cold spells or extreme weather, classical assumptions turned out to be too optimistic.

Another key insight: every single forecast must be stored permanently. Only the systematic archiving of predicted and later actual values enables regulatory traceability, true retrospective validation and operational improvement.

The economic order of magnitude of the reform is substantial. A cost-benefit analysis by Ofgem, the British regulator for the gas and electricity markets, puts the cumulative efficiency gains of the MHHS rollout for the British electricity market at GBP 1.6 to 4.5 billion over the period 2021 to 2045. At supplier level, industry estimates suggest that significantly improved load forecasting can lower balancing-energy costs by 8 to 15 percent. For a mid-sized British supplier with around 3 TWh annual volume, that corresponds to annual savings in the single-digit million-pound range.

The three most important lessons

1. Average accuracy can be dangerously misleading.
   The first dashboards looked excellent — until the models were broken down by time of day and customer segment. Forecasts were systematically off in precisely the most expensive hours. Modern forecasting systems must therefore be evaluated granularly: by settlement period, weather pattern, season and customer type.
2. Backtests must simulate real production conditions.
   A static backtest ignores drift, behavioural changes and regime shifts. That is why models today are retrained daily and validated using the so-called walk-forward method.
3. Data pipelines are just as critical as the models themselves.
   A faulty weather feed or a smart-meter issue on a daylight-saving-time day is enough to produce plausible but wrong forecasts. Modern systems are therefore monitored continuously: they automatically flag when forecast quality starts to drop and log every single forecast completely and traceably.

What can be transferred to Switzerland — and what cannot

Switzerland is not a copy of the British electricity market. The market structure is different, the balancing-group logic differs and quarter-hourly settlement has been established here for longer.

Nevertheless, key building blocks can be transferred directly:

• the separation of load, price and feed-in forecasts,
• walk-forward backtests,
• persisted forecast-vs-actual databases,
• AI agents for evaluating REMIT alerts and weather events,
• and hybrid ML/fundamentals models.

Other components, by contrast, require Swiss redevelopment. Particularly demanding are:

• alpine PV forecasts with microclimates,
• hydrology and reservoir-inflow models,
• pumped-storage optimisation,
• and balancing-group wrappers for Swissgrid and the European Target Model.

It is precisely hydropower that makes Switzerland unique. While Britain has virtually no relevant storage hydrology, Swiss hydropower is increasingly acting as Europe's «national battery».

Conclusion

Switzerland does not have to start from scratch on AI in the energy industry. Much has already been developed, tested and run in production internationally. The real opportunity now lies in adapting this experience intelligently and shaping it into a Swiss solution: pragmatic, regulatorily clean, and together with the utilities.

Because forecast quality will not just be an operational capability in the coming years. It will become a strategic competitive factor for the entire electricity industry.