AI for energy and utilities: from load forecasting to smart grid optimisation

The Dutch electricity grid is gridlocked. Grid congestion is blocking new connections, the balance between supply and demand shifts with every MWh of solar and wind generation, and the price curve on EPEX increasingly deviates from what classic forecasting predicts. Energy companies, grid operators, ESCOs and EV charging operators that want to stay afloat are deploying artificial intelligence for load forecasting, predictive maintenance on transformer stations, balancing services and EV charging orchestration. We build those AI solutions, tailored to the Dutch grid code, TenneT/Stedin/Liander/Enexis protocols and the reality of a grid that runs closer to its limits every day.

Load forecasting Smart grid optimisation Predictive maintenance EV charging orchestration Energy trading AI DERMS & ADMS
Discuss your AI case View use cases

Why AI and energy have become inseparable

Every second, the energy sector generates measurement data from millions of smart meters, SCADA systems, weather stations, EV chargers and home batteries. Those data volumes are only manageable with machine learning, and market margins can only be defended with models that predict faster than the competition.

In 2026, the Dutch electricity grid is structurally congestion-prone in six of the seven regional grid operator (RNB) regions. Liander, Stedin and Enexis publish weekly updated capacity maps on which entire provinces are coloured orange or red. At the same time, the share of solar and wind generation is growing at a pace that SCADA architecture from the coal and gas era was never designed to handle. Day-ahead prices on EPEX are negative every day between 10 a.m. and 4 p.m. and spike to well above €300/MWh in the evening peak. Classic forecasting models, which worked statically with seasonal patterns and historical profiles, have an unacceptable margin of error in this reality.

AI models trained on smart meter data, weather forecasts from KNMI or ECMWF, OCPP messages from charging stations and order book data from EPEX and Nord Pool can predict demand, supply and price minute by minute with solid grounding. That predictive power is worth money directly: for a BRP (Balance Responsible Party), halving the imbalance forecast error means a correspondingly lower imbalance cost line. For a grid operator, seeing a transformer failure coming three weeks ahead means an avoided emergency intervention. For an EV charging operator, smart orchestration of hundreds of charging sessions means the grid connection capacity doesn't need to be upgraded, an investment that would otherwise run into the millions.

Key areas where AI makes the difference in energy and utilities

From forecasting to fraud detection, the applications vary, but the common thread is the same everywhere: lower imbalance costs, higher utilisation of existing grid capacity and demonstrably less downtime.

📊

Load forecasting for BRPs and suppliers

Quarter-hourly supply and demand forecasting for the day-ahead, intraday and imbalance markets. Models combine E1/E2 meter data, ECMWF weather ensembles, calendar effects and macro indicators. The output can be used directly for your e-programme submission to TenneT, for bid strategy on EPEX and for optimising your BRP portfolio within the aFRR/mFRR/FCR balancing products.

⚡

Smart grid optimisation for DSOs

Real-time congestion management at medium- and high-voltage level via DERMS and ADMS integrations. AI models predict local congestion before it occurs, prioritise GOPACS bids and control flexible assets via OpenADR 2.0b. Includes interfaces to IEC 61850 substations and your existing SCADA environment.

🔧

Predictive maintenance on critical assets

Substations, switchgear, wind turbines and solar park inverters generate vibration, temperature and current data. Anomaly detection and remaining-useful-life models predict failures weeks before they occur. Maintenance shifts from reactive to planned: fewer emergency call-outs, less unplanned downtime and longer asset lifespans.

🔌

EV charging orchestration and V2G

Smart control of charging sessions via OCPP 2.0.1 and ISO 15118. Models optimise charging curves based on dynamic tariffs, local grid congestion, behind-the-meter PV generation and user preferences. Vehicle-to-grid integration in which EV batteries actively participate in aFRR or FCR-Day. Suitable for CPOs, fleet operators and utility owners of charging infrastructure.

📈

Energy trading AI for trading desks

Reinforcement learning agents for day-ahead bidding on EPEX and APX, intraday cyclic trading and cross-border arbitrage with Nord Pool. Models learn from order book dynamics, weather deviations and cross-commodity correlations. Hands-off trading within predefined risk frameworks, with a full audit trail for REMIT reporting and internal compliance.

🏠

Prosumer management and HEMS

Behind the connection: coordinated control of solar panels, home battery, heat pump and EV charger. Models optimise self-consumption, feed-in timing and participation in dynamic tariffs. For suppliers with dynamic contracts, a tool that helps customers earn more from their PV installation while improving portfolio balance.

💬

Customer service AI for consumption queries

A large share of inbound tickets at energy suppliers concern annual statements, deviating meter readings, feed-in settlement and house moves. LLM-based assistants retrieve current consumption data from your billing system, explain deviations clearly and handle routine tickets. Escalation to a human with full context handover, so no frustrating transfers.

🛡️

Fraud detection on meter readings

Unexplained drops in consumption, reversal of supply direction, tampering with smart meters: anomaly detection models identify suspicious patterns in smart meter data. Useful for suppliers, DSOs and metering companies to reduce grid losses and investigate non-technical losses in a targeted way.

🌬️

Wind and solar yield forecasting

Probabilistic yield forecasting per asset on a quarter-hourly basis. Models combine NWP data (HARMONIE-AROME, ECMWF), turbine SCADA and historical curtailment data. Indispensable for BRPs with large PPA portfolios and for operators who want to optimise their day-ahead nomination within the Dutch imbalance regime.

How Appfront builds AI for energy and utilities

We don't build a generic SaaS platform that you have to configure around your grid setup yourself. Our approach is hands-on: we dive into your SCADA architecture, energy management systems and trading stack, and build AI components that integrate directly with the reality of your operation. Whether that's a Siemens Spectrum Power installation, a GE PowerOn, a proprietary DERMS platform, or a combination with cloud billing such as Stedin Klant or the Greenchoice stack.

Every energy organisation starts from a different position. A large grid operator with thousands of substations and an in-house data lake has different priorities from a new ESCO rolling out a charging network on an industrial estate, or a small BRP that wants to halve its imbalance costs with better forecasts. We tailor the AI architecture to your specific reality: which metering and SCADA data is actually available and reliable, which regulatory requirements from the ACM and the Dutch Grid Code apply, and where the biggest impact lies on imbalance costs, OPEX or customer satisfaction.

Our way of working is iterative and deliberately careful with production environments. We start with a proof of concept on a single, well-defined use case, such as load forecasting for one BRP portfolio or predictive maintenance on one type of substation, and expand once the model results are convincing enough for a production rollout. No big-bang implementation in critical infrastructure, but a validated rollout with measurable KPIs: avoided imbalance costs, avoided emergency interventions, avoided curtailment.

From first conversation to AI in production on your grid operation

Our approach to AI projects in the energy sector follows four phases. Each phase delivers a concrete result: no months of analysis without output, but interim milestones that you can test against your own operational reality.

Data assessment

We take stock of which data is available: SCADA historian, smart meter data, OCPP logs, weather data, EPEX history. We assess quality, latency and reliability, and determine which AI application is the most feasible technically and economically.

Proof of concept

Within a few weeks we build a working prototype of the chosen use case: a forecasting model trained on your own portfolio, anomaly detection on one type of substation, or an orchestration engine for one charging hub. Concrete and testable.

Integration and production

The validated model is connected to your existing systems: SCADA, ADMS, billing platform, trading stack. We build IEC 61850 or OpenADR integrations, dashboards for operators and monitoring for your data team.

Monitoring and adjustment

AI models go stale when the market or the grid shifts. We monitor model performance, detect drift, retrain on new data and adjust based on changing grid conditions, regulations and feedback from your operations team.

Not sure about a large project yet?

Test your idea first: a working prototype in 1 day

With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.

View OneDayBuild →

Technology we use in energy projects

The choice of technology depends on the use case, the latency requirements and the existing IT/OT architecture. For load forecasting we work with gradient boosting (LightGBM, XGBoost) and sequence models such as N-BEATS or Temporal Fusion Transformers. For predictive maintenance we use time-series anomaly detection, autoencoders and survival models for remaining useful life. For energy trading AI we build reinforcement learning agents within pre-calibrated risk frameworks and strict safety rails.

We choose technology based on proven results in the utility sector, not on hype. Where a simple regression model is enough for day-ahead, we don't build deep transformers. Where an LLM is needed for customer service or data explainability, we make sure the model answers in a controlled way — no hallucinations about tariffs, contract terms or meter readings. And for anything involving SCADA and critical infrastructure: strict separation between IT and OT environments, in line with IEC 62443.

Python LightGBM XGBoost N-BEATS Temporal Fusion Transformer PyTorch TensorFlow Apache Kafka TimescaleDB Apache Airflow FastAPI OpenADR 2.0b OCPP 2.0.1 IEC 61850 MLflow Docker / Kubernetes

Energy data, the Dutch Grid Code and compliance: what you need to know

Working with data from the Dutch electricity grid means working within a dense body of regulation: GDPR, the Electricity Act 1998, the Dutch Grid Code (Netcode Elektriciteit), the Information Code (Informatiecode), ACM decisions and — for trading activities — REMIT and MiFID II. We build AI solutions that stay within these frameworks.

Grid Code and Information Code

Access to metering and allocation data is governed by the Information Code (Informatiecode). Our pipelines respect market roles (BRP, supplier, DSO, metering company) and the associated data minimisation. No unauthorised data exchange between market roles, and no cross-contamination between regulated and commercial activities — a point on which the ACM enforces strictly.

GDPR for smart meter data

Smart meter data is personal data. We implement pseudonymisation at connection EAN level, explicit legal bases for processing and clear data processing agreements. Training data for portfolio-wide models is aggregated or treated with differential privacy wherever possible.

REMIT and trading transparency

Energy trading AI operates within REMIT's transparency requirements. Model decisions are logged, including inputs, predictions and chosen actions — so the ACM, the AFM or an internal compliance officer can always reconstruct why a particular bidding strategy was chosen at a given moment.

NIS2 and cybersecurity for OT

Under NIS2, energy suppliers, DSOs and large producers count as essential entities. AI models that use SCADA or ADMS data operate within IEC 62443 zoning, with audit logging, MFA access and strict separation between model training and production environments. No direct cloud uplinks from OT zones without a hardened gateway.

Hosting within the NL/EU

Sensitive metering, SCADA and trading data never leaves the Netherlands or the EU. We host AI models and data lakes with Dutch cloud providers, on-premise or within your existing Azure/AWS tenancy in an EU region. No data export to a US cloud without your explicit choice and an accompanying DPIA.

Explainable AI for regulators

A model that makes decisions about balancing, congestion management or customer invoices must be explainable. We deliver explainability through SHAP values, feature attribution and clear decision rules — usable for your audits, for ACM enquiries and for the trust of your own operations team.

Concrete scenarios from Dutch energy practice

AI in energy is not a vision of the distant future — these are realistic applications that can be built for your organisation now, with the current state of technology and data availability.

BRP portfolio with dynamic forecasting

A Balance Responsible Party with a mixed portfolio of wind, solar and gas assets uses an ensemble model that combines ECMWF weather data, turbine SCADA and historical imbalance patterns. The model delivers 192 quarter-hourly values for the next 48 hours, updated every hour. The output feeds directly into the e-programme engine that submits nominations to TenneT and places intraday bids on EPEX. Result: substantially lower imbalance costs and better use of flexible assets.

Grid operator with predictive maintenance on MV substations

A regional grid operator collects vibration, temperature and load data from medium-voltage substations. An anomaly detection pipeline runs daily and flags abnormal patterns weeks before a potential failure. The maintenance team plans targeted inspections and replacements instead of reacting to failures. Result: lower SAIDI/SAIFI, better asset utilisation and a more targeted maintenance budget.

EV charging operator with smart orchestration at charging hubs

A CPO with charging hubs on industrial estates uses an orchestration engine that adjusts charging curves in real time based on local grid congestion, dynamic tariffs and user preferences. Communication runs via OCPP 2.0.1, with congestion signals via OpenADR. Result: existing connection capacity is sufficient despite growing charging volumes, avoiding significant grid reinforcement investment.

Energy supplier with dynamic tariffs and HEMS integration

A supplier with a dynamic contract integrates a HEMS link that controls the home battery, heat pump and solar panels based on hourly prices and local grid congestion signals. AI orchestration optimises self-consumption and feed-in revenue for the customer, while the supplier improves its portfolio balance. Customers see lower annual bills, while the supplier sees lower imbalance costs and higher retention.

ESCO with flex aggregation for balancing services

An ESCO bundles heat pumps, cold stores and industrial batteries into a virtual power plant and offers aFRR and mFRR capacity to TenneT. Every 15 minutes, AI determines which assets are available, how much they can deliver and at what cost. Bidding on the balancing market happens automatically, within predefined risk frameworks and with a full audit trail.

Utility supplier with customer service AI for consumption queries

A supplier receives seasonal peaks in tickets around annual statements and balance corrections. An LLM assistant with a direct integration to billing and EDSN data answers consumption queries, explains anomalous meter readings and flags potential meter problems for referral back to the metering company. Result: shorter handling times, a higher first-time-fix rate and freed-up capacity for more complex issues.

Frequently asked questions about AI for energy and utilities

Does your AI work with our existing SCADA and ADMS?
Yes. We integrate with common SCADA and ADMS platforms via standard protocols (IEC 61850, IEC 60870-5-104, OPC UA, MQTT). For grid operators, we build interfaces that stay within IEC 62443 zoning and respect the separation between IT and OT domains. For SCADA historians, we work with read-only replicas or dedicated aggregates, so production environments remain unaffected.
How much can we structurally save on imbalance costs with better forecasting?
That depends heavily on your portfolio composition, your current forecasting stack and market conditions. For portfolios with substantial wind and solar components, we generally see modern ML models noticeably reduce the imbalance forecast error compared with classical models. We make no hard savings promises upfront. We start with a proof of concept on your own historical data and show you what the difference is in concrete terms for your portfolio.
Can you integrate with TenneT, Stedin, Liander and Enexis protocols?
Yes. We have experience with the market-role-specific integrations: e-programme submission to TenneT, GOPACS bidding for congestion management, EDSN metering data for suppliers, and the regional grid operator (RNB) portals for capacity requests. For each project, we assess which integrations are relevant and build them in line with the applicable interface specifications.
Does EV charging orchestration also work with OCPP 1.6, and not just 2.0.1?
Yes. We support both OCPP 1.6 and 2.0.1, with a gradual migration path towards 2.0.1 for new installations. V2G applications require ISO 15118 in combination with OCPP 2.0.1. We also help you choose between local smart charging via a charge station management system and cloud-based orchestration.
Can an AI model trade autonomously on EPEX or Nord Pool?
Under REMIT it can, provided there is a full audit trail and it operates within defined risk frameworks. We always implement safety rails: maximum position size, hard stop-loss limits, human-in-the-loop for exceptions and extensive logging. Additional requirements apply to MiFID II-regulated activities, and we factor these into the design.
How do you handle the data quality of smart meters and SCADA?
In our experience, data quality is the biggest obstacle to AI in utilities. We start every project with a data assessment: missing values, clock deviations, duplicate measurement series, sensor drift. Only once the data quality is in order do we start modelling. For production, we build automated data validation steps that reject input if quality drops below the threshold.
How long does it take for an AI solution for energy to go into production?
A proof of concept is usually ready within a few weeks. The lead time to production depends heavily on integration requirements: SCADA integrations, NIS2 compliance and internal change management processes in particular take time. We work in iterative sprints and deliver interim milestones, so you see results along the way and can adjust course.
Do you also work for smaller ESCOs and CPOs, or only for large players?
Both. Our approach scales: a large grid operator with a data lake and in-house data engineers gets a different solution from an ESCO with a few dozen assets. The starting point remains the same (a concrete use case, a proof of concept, validated rollout), but the architecture and cost scale differ considerably.

Want to use AI in your energy or utility organisation?

Discuss your case with us. We analyse where AI delivers the most value within your grid operations or trading stack, with no obligation.

Schedule a call

Edit content