Distribution grid congestion management & redispatch optimisation (DSO/TSO)

Heat pumps, wallboxes and rooftop PV push low- and medium-voltage grids to their limits long before copper and transformers can be expanded. Grid operators combine state estimation from smart meters, congestion forecasts and power-flow-aware optimisation to decide which flexibility to call, which generator to redispatch and which controllable load to throttle, and when. The aim is to defer or avoid grid expansion, reduce curtailment and keep supply secure, with the operator always keeping the final control authority.

Problem & value

• Today: the low-voltage grid is largely unobserved, planning relies on worst-case assumptions and standard load profiles, and congestion is handled by static rules, manual switching or conservative connection limits. Grid-connection queues for PV, storage and charging infrastructure grow in many EU countries.
• AI value: better observability (state estimation with few sensors), probabilistic congestion forecasts days to minutes ahead, faster optimisation of redispatch and flexibility calls under AC power-flow constraints, and automated grid-connection checks.
• Regulatory driver (Germany): Redispatch 2.0 integrates small distributed generators into congestion management; section 14a EnWG lets operators curtail controllable consumption devices (heat pumps, wallboxes, storage) in exchange for reduced grid fees [1][2]. At EU level, Regulation (EU) 2019/943 and the Electricity Directive (EU) 2019/944 require market-based procurement of flexibility by DSOs where economically efficient [3].
• Benefits: industry bodies such as E.DSO describe AI as a way to defer grid investment and improve operations, but no neutral, comparable benefit figures exist across operators; measure on your own grid [4].

Variants

• DSO congestion forecasting and flexibility procurement: forecast overloads on MV/LV feeders and transformers, call flexibility via a local market or platform.
• DSO controllable-load management (section 14a EnWG): dimming of wallboxes and heat pumps by the grid operator, with grid-status traffic-light signals.
• TSO redispatch optimisation: security-constrained optimal power flow to select redispatch and countertrading measures, cost-optimal and fair across plants.
• Dynamic line rating (DLR): ratings from weather and conductor sensors or models raise overhead-line capacity within safe limits.
• Grid-connection check and expansion planning: automated hosting-capacity calculation for connection requests, scenario-based network planning.
• Industrial site (manufacturing): plants with large controllable loads provide flexibility to the operator (see OPT-011, OPT-007).

How it works

1. grid model (GIS/NIS topology) + measurements (smart meters, SCADA, PMU) + weather
2. state estimation for unobserved LV/MV grid (physics + ML pseudo-measurements)
3. load / PV / wind forecasting (probabilistic, per feeder)
4. congestion forecast (power-flow simulation, N-1 check)
5. optimisation (OPF / MILP, flexibility bids, redispatch cost minimisation)
6. operator review and approval (control-room dispatcher)
7. control signals (SCADA, flex platform, EMS/HEMS gateways) + audit log

• State estimation: weighted least squares combined with learned pseudo-measurements or graph neural networks (graph-ml) to infer voltages and flows where there are few sensors.
• Forecasting: probabilistic time-series models, for example gradient boosting or transformer-based time-series foundation models, fed with weather and smart-meter aggregates (see PRED-007).
• Optimisation: AC or linearised optimal power flow solved with MILP/NLP solvers; reinforcement learning is used mostly as a warm-start or in simulation, not as the sole safety-critical controller.
• Human-in-the-loop: dispatchers approve or override proposed measures; hard safety limits remain in deterministic protection and SCADA logic outside the AI.

Data preparation

Needed: grid topology and asset data (GIS, grid-planning tools), real-time and historic measurements (SCADA, smart meters), generation and load forecasts, redispatch and market data, and the counterparty data of flexibility providers. It lives in GIS, SCADA/DMS, market-communication systems and many Excel files.

Typical gaps: inconsistent asset IDs between GIS and SCADA, outdated topology, missing or low-resolution measurements in low-voltage grids, data silos between grid and market departments, restricted access rights for operational data, and smart-meter data subject to strict legal limits.

Preparation steps: (1) data inventory and ownership (PREP-001, PREP-004); (2) extraction from SCADA, GIS and market systems (PREP-005); (3) harmonise asset keys and units, build a tidy time-series store (PREP-011, PREP-012); (4) pseudonymise or aggregate meter data before modelling (PREP-022, PREP-025); (5) assemble a validation set of past congestion events with grid planners (PREP-018). Tools: time-series database, pandas, pandapower models.

Rough share of effort: 40-50 % of the project, the highest share of all steps, mainly for harmonisation and legal clearance. See GUIDE-003.

Implementation paths

Which path for whom

• enterprise: Path B, because congestion management is core to grid operators and must be integrated with own grid models, SCADA/DMS and market processes, which standard products only partly cover; Path A modules (e.g. from the existing grid-control vendor) are a sensible building block.
• developer: Path C, because open grid tools (pandapower, PyPSA) and solvers run on HW1 and allow transparent, auditable logic.
• public: Path C, because municipal and regional grid operators fall under critical-infrastructure and sovereignty requirements and keep sensitive grid data on their own or German sovereign infrastructure.

Path A — Buy (off-the-shelf)

• Grid-planning and operation software: Envelio (EU, Germany, Intelligent Grid Platform), Siemens Gridscale X (EU), Hitachi Energy Network Manager / ADMS, GE Vernova GridOS, Schneider Electric EcoStruxure ADMS, Landis+Gyr and Itron for metering analytics, PSI Software (EU) for control systems, Smart Wires for grid-enhancing hardware.
• Flexibility platforms: GOPACS (Netherlands, joint TSO/DSO platform), Piclo Flex, Enera, Nodes (Nordic, EU), Equigy for crowd balancing.
• DLR vendors: Heimdall Power, LineVision, Ampacimon (EU).
• Choose A when: you need production-grade, certified integration with SCADA/ADMS and vendor support; most DSOs and Stadtwerke fall here.

Path B — Build on cloud APIs

• Forecasting and analytics on EU cloud regions, using managed time-series services or time-series foundation models via API, with data kept under a processing agreement. The network model and control commands stay on-premises.
• LLMs are an assistant layer only: summarising congestion reports, answering planners' questions over network data, generating regulatory documentation (see OPT-018 for LLM-to-solver modelling).
• Choose B when: you have a data team and want custom analytics next to a vendor ADMS, but control actuation remains in certified systems.

Path C — Self-host (open source)

• Power-flow and grid analysis: pandapower, PyPSA, PowerModels.jl, GridCal, OpenDSS; optimisation: HiGHS, SCIP, Pyomo, OR-Tools, plus Gurobi or CPLEX commercially.
• Forecasting: LightGBM, Darts, open time-series foundation models such as Chronos or TimesFM; streaming: Kafka plus a time-series database.
• Standards and platform: the LF Energy ecosystem (for example OpenSTEF for forecasting, SEAPATH for substations), IEC 61970 CIM data models.
• Hardware: HW1 (CPU servers) is sufficient for most forecasting and OPF; GPUs matter only for GNN training.

Path D — Custom training

Worthwhile for operator-specific forecasting models and GNN-based state estimation when you have years of smart-meter and SCADA history. Train on historical measurements plus simulated grid scenarios, validate against held-out N-1 events, and keep a physical-model fallback.

State of the art (2026)

• Hybrid physics plus ML is the norm: ML provides forecasts and pseudo-measurements, physics-based power flow and optimisation provide feasibility guarantees.
• GNN and learned OPF surrogates speed up power-flow solutions by orders of magnitude in research, but guarantees on constraint satisfaction and rare topologies remain open; they are used as warm starts or screening tools.
• TSO/DSO coordination and flexibility markets (GOPACS, local flex markets) are expanding; standardised data exchange through CIM and ENTSO-E processes is the bottleneck.
• Limits: poor topology data quality, missing LV measurements, forecast errors during extreme weather, gaming of flexibility markets (inc-dec behaviour), and cyber-security of control paths. Rare events are underrepresented in training data.

Shared resources & combinations

• Same stack (time-series, optimization-solver): OPT-011 (battery dispatch), OPT-021 (EV smart charging), OPT-032 (home energy management), OPT-039 (renewable marketing), OPT-007 (building energy) deliver the flexibility the operator calls.
• Upstream: PRED-007 (load and renewable forecasting), ANOM-008 (grid anomaly and theft detection, shares the smart-meter data pipeline).
• Downstream: OPT-023 (district heating dispatch) for sector coupling.
• Shared solver library and time-series platform serve all of them; grid models are reusable for planning and operations.

Effort, cost & KPIs

• Path A: 6 to 18 months including ADMS and metering integration; licence plus integration cost dominates.
• Path B/C: 6 to 12 months for forecasting and decision-support pilots on one grid area; actuation integration adds time.
• KPIs: forecast error (MAE/CRPS) per feeder, avoided curtailed energy, redispatch cost, number of overloads and N-1 violations, flexibility cost per MWh, grid-connection processing time, share of operator proposals accepted.

People & change

Involve management and the grid-operations lead first; grid planning, control-room staff and market/flexibility teams are the core users. IT and especially the information-security officer must be in early, as the system touches critical infrastructure (GUIDE-013). Data owners of SCADA, GIS and metering give access; legal/compliance clarifies regulatory duties (EnWG, redispatch rules); the DPO is needed because smart-meter data is personal.

Works council: usually no for the optimisation itself, since it works on grid and market data, not on employee behaviour. Still inform the works council in good time under §90(1) no. 3 BetrVG, and §87(1) no. 6 applies as soon as control-room actions are logged and evaluated per person. Public operators involve the Personalrat. See GUIDE-004. DPIA: likely if household-level meter data is used; prefer aggregation (GUIDE-005).

AI literacy: train control-room staff on forecast uncertainty and failure modes (GUIDE-006), consistent with AI Act Art. 4 (measures to support literacy). Pilot with one grid area in shadow mode, where the system only recommends. Typical concerns: loss of operational control and liability for wrong switching. Address them by having operators approve every measure, keeping manual fallback and documenting decisions (GUIDE-008, GUIDE-009).

Risks, regulation & compliance

• EU AI Act: AI systems used as safety components in the management and operation of critical digital infrastructure, road traffic or the supply of water, gas, heating or electricity fall under Annex III point 2 and are high-risk [5]. Pure forecasting or planning tools that do not act as safety components may be minimal risk. Hence classification is "depends"; assess each deployment. No Art. 5 prohibited practice is involved.
• Other EU law: NIS2 Directive (energy sector is an essential entity) and the Critical Entities Resilience Directive for cybersecurity and resilience; Cyber Resilience Act for products with digital elements; national energy law (EnWG, Redispatch 2.0 rules, section 14a) and network codes [1][2][3].
• GDPR: smart-meter data at household level is personal data; aggregate to feeder level where possible, minimise granularity (15-minute data is sensitive), and base processing on legal grounds in the metering law (Messstellenbetriebsgesetz) and DPIAs.
• Risks and mitigations: wrong curtailment of households or plants (fairness and compensation rules), unreliable forecasts (probabilistic bounds, fallback to conservative limits), cyber attack on control paths (segmented networks, signed commands), model drift (monitoring), opaque decisions (audit log of inputs, constraints and operator decisions).

Getting started

1. Inventory data: topology quality in the GIS/NIS, smart-meter coverage, SCADA channels, historical congestion events.
2. Build a feeder-level probabilistic load and PV forecast and backtest it against known overload events.
3. Run power-flow simulations with pandapower on one grid area to identify likely congestion hotspots.
4. Pilot a decision-support dashboard (no automatic control) and compare proposals with dispatcher decisions.
5. Classify the system under the AI Act and NIS2 before any step toward automated actuation.

Sources

1. Bundesnetzagentur — Energy portal (Redispatch 2.0) — https://www.bundesnetzagentur.de/DE/Vportal/Energie/start.html
2. Bundesnetzagentur — Section 14a EnWG, steuerbare Verbrauchseinrichtungen — https://www.bundesnetzagentur.de/DE/Vportal/Energie/SteuerbareVBE/start.html
3. Regulation (EU) 2019/943 on the internal market for electricity — https://eur-lex.europa.eu/eli/reg/2019/943/oj
4. E.DSO — European Distribution System Operators — https://www.edsoforsmartgrids.eu/
5. Regulation (EU) 2024/1689 (AI Act), Annex III — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. pandapower — https://github.com/e2nIEE/pandapower
7. GOPACS — https://www.gopacs.eu/
8. LF Energy OpenSTEF — https://github.com/OpenSTEF/openstef