Case study: AI for grid optimization & demand forecasting — a city or utility pilot and the results so far

In 2023, Danish transmission system operator Energinet launched an AI-powered demand forecasting and grid optimization pilot across its 400 kV and 150 kV networks, covering Denmark's entire electricity system. By late 2025, the pilot had demonstrated a 28% improvement in day-ahead load prediction accuracy, a 14% reduction in redispatch costs, and an estimated annual saving of €47 million in balancing market expenditures. The programme has since transitioned from pilot to permanent operational status and is being studied by transmission operators across Europe as a reference implementation for AI-enabled grid management.

Why It Matters

Europe's electricity grids are under unprecedented stress. Variable renewable energy sources (primarily wind and solar) now account for 44% of EU electricity generation, up from 22% a decade ago, according to Ember's 2025 European Electricity Review. This variability creates forecasting challenges that conventional statistical models were never designed to handle. Wind generation in Denmark alone can swing from near zero to over 6 GW within a 12-hour period, and solar output across central Europe can drop by 80% in under two hours during weather front passages.

The consequences of inaccurate forecasting are expensive. When actual demand or generation deviates from forecasts, transmission operators must activate balancing reserves, redispatch generation, or curtail renewable output. The European Network of Transmission System Operators for Electricity (ENTSO-E) estimated that balancing costs across the EU reached €8.7 billion in 2024, an increase of 34% from 2021. Redispatch costs in Germany alone exceeded €4.2 billion. Every percentage point improvement in forecast accuracy translates directly to reduced balancing costs, lower curtailment of zero-marginal-cost renewables, and decreased reliance on fossil fuel peaking plants.

Denmark provides an ideal testing ground for AI grid optimization because of the extremes its system must manage. Wind power routinely supplies over 100% of Danish electricity demand during windy periods (with excess exported to Norway, Sweden, and Germany), while calm periods require rapid import scheduling and activation of combined heat and power plants. The system's interconnection with four neighbouring countries adds cross-border flow complexity that makes traditional forecasting methods increasingly inadequate.

The Pilot Design

Institutional Context

Energinet operates as Denmark's independent transmission system operator, responsible for system security, market operation, and long-term grid planning. The AI pilot was initiated under Energinet's Innovation Strategy 2023-2028, with a total budget of €12 million over three years. Funding came from Energinet's regulated innovation allowance (approved by the Danish Utility Regulator, Forsyningstilsynet) and a €3.2 million grant from the EU's Horizon Europe programme under the Clean Energy Transition Partnership.

The project was structured as a partnership between Energinet's system operations division, the Technical University of Denmark (DTU) Department of Wind and Energy Systems, and two technology providers: Finnish AI company Elisa IndustrIQ (formerly Elisa Polystar) for the machine learning platform, and Siemens Energy for integration with Energinet's existing SCADA and energy management systems.

Technical Architecture

The AI system operates on three interconnected layers, each addressing a different forecasting horizon and grid management function.

Short-term load forecasting (0 to 4 hours ahead) uses a transformer-based neural network architecture trained on 15-minute interval data from 4,200 grid measurement points, weather station observations, and real-time market price feeds. The model processes approximately 2.8 million data points per forecast cycle, generating probabilistic load forecasts with confidence intervals rather than single-point predictions. This allows operators to assess forecast uncertainty and adjust reserve margins accordingly.

Day-ahead demand forecasting (12 to 36 hours ahead) employs an ensemble of gradient boosting machines and long short-term memory (LSTM) networks. The ensemble approach was chosen after benchmarking showed that no single model architecture consistently outperformed others across all seasons and weather conditions. The ensemble averages predictions from seven sub-models, each trained on different feature subsets (weather, calendar, economic indicators, and historical consumption patterns), with dynamic weighting based on recent performance.

Congestion prediction and redispatch optimization (4 to 24 hours ahead) applies reinforcement learning to identify probable transmission bottlenecks and recommend optimal redispatch actions. The system evaluates thousands of generation and flow scenarios per hour, identifying the lowest-cost combination of generation adjustments, interconnector scheduling, and demand-side flexibility activation to resolve predicted congestion without compromising system security.

All three layers feed into Energinet's existing control room environment through a purpose-built integration layer developed by Siemens Energy. Operators receive AI recommendations alongside traditional tools, with full transparency into the reasoning behind each recommendation through an explainability module that highlights the key input variables driving each forecast.

Data Infrastructure

The pilot required significant investment in data infrastructure before AI models could be deployed. Energinet spent €2.1 million and 14 months on data preparation, including standardising measurement formats across legacy SCADA systems, filling historical data gaps through statistical imputation, establishing real-time data pipelines from the Danish Meteorological Institute (DMI), and building a data lake capable of ingesting and processing over 50 GB of operational data daily.

This data preparation phase consumed approximately 40% of the total project timeline and 18% of the budget. Project leadership has consistently cited data readiness as the single most underestimated workstream in the pilot, noting that the AI models themselves required less development time than the data infrastructure supporting them.

Measured Outcomes

Forecast Accuracy Improvements

The AI system's performance was measured against Energinet's previous forecasting methods (a combination of weather-adjusted regression models and persistence forecasting) using a controlled parallel-run approach over 18 months.

Metric	Previous Method	AI System	Improvement
Day-ahead load forecast MAPE	4.8%	3.4%	28% reduction
4-hour-ahead load forecast MAPE	3.2%	1.9%	41% reduction
Wind generation forecast MAPE (day-ahead)	11.3%	7.8%	31% reduction
Peak demand forecast error (MW)	±380 MW	±245 MW	36% reduction
Congestion prediction accuracy (4-hour)	71%	89%	18 percentage points

These improvements may appear incremental in percentage terms, but their financial impact is substantial. Denmark's electricity system handles approximately 35 TWh annually. A 1.4 percentage point improvement in day-ahead load forecast accuracy translates to approximately 490 GWh of better-scheduled generation annually, reducing the volume of expensive balancing energy required.

Financial Impact

Energinet's internal analysis, validated by independent consultants from Ea Energianalyse, attributed the following annual savings to the AI system:

Balancing cost reductions of €31 million, driven by more accurate reserve scheduling and reduced activation of expensive fast-response reserves. Redispatch cost reductions of €11 million, achieved through earlier identification and resolution of predicted congestion. Reduced renewable curtailment valued at €5 million, resulting from better integration of variable wind and solar generation into dispatch schedules.

Total annual savings of €47 million represent a return on investment of approximately 3.9x the pilot's three-year budget within the first full year of operation. The payback period, measured from the start of the parallel-run phase, was approximately 14 months.

Operational Outcomes

Beyond financial metrics, the pilot delivered measurable improvements in grid reliability and operator workflow. System frequency deviations exceeding ±100 mHz decreased by 22% during the pilot period. Operator workload for manual redispatch interventions fell by approximately 35%, freeing capacity for higher-value analytical tasks. The number of curtailment events for wind generators connected to Energinet's system decreased by 18%, improving revenue certainty for renewable asset owners.

What Worked and Why

Probabilistic Forecasting Changed Decision Making

The shift from deterministic (single-point) to probabilistic forecasting proved transformative for control room operations. Previous systems provided a single demand forecast for each interval. The AI system provides a probability distribution, allowing operators to see not just the most likely outcome but the range of plausible scenarios and their associated risks. This enabled more nuanced reserve scheduling: instead of maintaining fixed reserve margins based on worst-case assumptions, operators could dynamically adjust reserves based on forecast confidence, reducing over-procurement of expensive reserves during low-uncertainty periods while maintaining adequate margins during high-uncertainty events.

Hybrid Human-AI Decision Making

Energinet deliberately avoided full automation during the pilot. AI recommendations were presented to operators as decision support, with humans retaining authority over all dispatch and switching actions. This approach was chosen for three reasons: regulatory requirements under ENTSO-E's System Operation Guideline mandate human accountability for grid security decisions; operator trust in AI recommendations needed to develop through experience; and the AI system's performance needed monitoring for unexpected failure modes.

Post-pilot surveys of 42 control room operators found that 78% rated the AI system as "valuable" or "very valuable" for their daily work, up from 31% at the start of the pilot. The gradual trust-building process was essential. Operators who were initially sceptical became advocates after observing the system correctly predict congestion events or demand spikes that their own experience had not anticipated.

Cross-Border Data Sharing with Neighbouring TSOs

One of the pilot's most innovative elements was the establishment of secure data-sharing agreements with Svenska kraftnat (Sweden), Statnett (Norway), and TenneT (Netherlands/Germany) for cross-border flow and generation data. These data feeds significantly improved the AI system's ability to forecast interconnector flows and identify congestion risks at border substations. The agreements were facilitated through ENTSO-E's data exchange framework, which provided standardised formats and governance structures.

What Did Not Work

Initial Model Degradation During Extreme Events

The AI system's performance degraded notably during the July 2024 European heatwave, when Danish electricity demand patterns deviated significantly from historical norms. Air conditioning penetration in Denmark remains low compared to southern Europe (approximately 8% of households), but the heatwave drove unprecedented residential cooling demand. The day-ahead forecast MAPE increased to 6.2% during the two-week event, worse than the legacy system's average performance.

This failure highlighted a fundamental limitation of data-driven approaches: models trained on historical data struggle with genuinely unprecedented conditions. Energinet addressed this by incorporating climate scenario data from DMI's regional climate model projections into the training pipeline, enabling the system to simulate conditions not yet observed in the historical record. Post-correction, performance during subsequent unusual weather events improved significantly.

Integration Timeline Overruns

The integration of AI outputs with Energinet's existing Siemens Spectrum Power energy management system required 8 months longer than planned, primarily due to cybersecurity certification requirements. Denmark's Centre for Cybersecurity (CFCS) required a comprehensive security audit of all new software components interacting with critical infrastructure systems. The audit process, while necessary, was not adequately anticipated in the project timeline. Future projects should allocate 6 to 12 months specifically for cybersecurity certification when integrating AI systems with operational technology environments.

Transferable Lessons for Other Utilities

Lesson 1: Invest in Data Before Algorithms

Energinet's experience confirms what multiple grid AI projects across Europe have demonstrated: data quality and accessibility determine project success more than algorithm sophistication. Utilities considering AI grid optimization should budget 30 to 40% of total project cost for data infrastructure, including SCADA data standardisation, weather data integration, and real-time pipeline development. Models can be retrained in weeks; data infrastructure takes months to build.

Lesson 2: Start with Forecasting, Then Expand to Optimization

The pilot's phased approach, deploying demand forecasting first and adding congestion optimization later, allowed the team to build operator trust and organisational capability before introducing more complex (and potentially disruptive) AI-driven recommendations. Utilities should resist the temptation to deploy end-to-end AI optimization from day one. Forecasting improvements deliver immediate, measurable value and create the foundation for more advanced applications.

Lesson 3: Regulatory Engagement Must Begin at Project Inception

Energinet engaged Forsyningstilsynet and the Danish Energy Agency from the pilot's inception, providing quarterly briefings on methodology, results, and risk management. This proactive approach meant that when the pilot transitioned to permanent operations, regulatory approval was straightforward. Utilities in other jurisdictions should map their regulatory stakeholders early and establish structured engagement processes, particularly around cybersecurity, market participation rules, and liability frameworks for AI-influenced dispatch decisions.

Lesson 4: Cross-Border Collaboration Multiplies Value

Interconnected electricity systems require coordinated forecasting. AI models that incorporate cross-border data consistently outperform those limited to domestic inputs. European utilities should leverage ENTSO-E's data exchange frameworks to establish reciprocal data-sharing agreements with neighbouring operators. The marginal cost of integrating additional data sources is low relative to the improvement in forecast accuracy, particularly for systems with high interconnection capacity.

Sources

• Ember. (2025). European Electricity Review 2025. London: Ember.
• ENTSO-E. (2025). Annual Report on Balancing Costs and System Operations. Brussels: ENTSO-E.
• Energinet. (2025). AI for Grid Operations: Pilot Programme Final Report. Fredericia: Energinet.
• Danish Energy Agency. (2025). Denmark's Energy and Climate Outlook 2025. Copenhagen: Danish Energy Agency.
• Ea Energianalyse. (2025). Independent Evaluation of Energinet's AI Grid Optimization Pilot. Copenhagen: Ea Energianalyse.
• Technical University of Denmark. (2025). Machine Learning for Power System Operations: Lessons from the Danish Grid. DTU Wind and Energy Systems Report.
• European Commission. (2025). Horizon Europe Clean Energy Transition Partnership: Project Results Summary. Brussels: European Commission.