Load Forecasting

Predicting future electricity demand, used at every layer of the power system — from real-time grid control to 30-year network investment planning. The forecast horizon determines which decisions can be made; the method determines whether those decisions are reliable. Without load forecasting, neither market participants nor grid operators can act efficiently.

Why forecasting matters for flexibility markets

Load forecasting is the foundational data layer for operational flexibility market participation. It appears at every stage of the flexibility service chain:

• DSO day-ahead procurement — DSOs procure flexibility based on a forecast that a specific grid segment will exceed its limit during a specific window. In CoordiNet, SWITCH, sthlmflex, and all operational Swedish LFMs, the procurement call is triggered by the DSO’s local load forecast. Without it, the DSO cannot know when to buy or how much. For availability-based products (LFM-h/p), the DSO must forecast which hours are congestion-risk in order to define the activation window in the product. DSO forecasting is therefore what turns a grid need into a market order.
• BRP plan submission — Balance Responsible Parties must submit consumption forecasts to Svenska kraftnät the day ahead; forecast error drives imbalance settlement charges under the Nordic enpris settlement model
• FSP bidding — Flexibility Service Providers cannot construct a bid (quantity, activation window) without knowing baseline consumption; the bid is the difference between baseline and committed curtailment
• DR scheduling — aggregators must predict when resources are available to respond; heat pump flexibility depends on weather/thermal state; EV flexibility depends on connection and departure time
• DSO capacity planning — grid operators must forecast local peak demand to decide whether to procure flexibility or reinforce the grid; DNDP load scenarios are built on MTLF/LTLF models

This places load forecasting as a prerequisite for the entire explicit flexibility market architecture on both sides of the trade — not a technical detail, but an enabling condition. Unknown demand = no market. (Source - Load Forecasting Methods Survey (2025))

But good STLF for a flexibility market is not the same as low MAPE: the forecast is a financial instrument defined by the market’s settlement rules, so “good” means accurate where errors are expensive, manipulation-resistant where money is at stake, and current at gate closure. For the full argument and the achievability levers, see STLF for Flexibility Markets — What Counts as Good and How to Achieve It.

Four forecasting horizons

The STLF tier is the operational market participation layer. MTLF and LTLF are used in Swedish regulation primarily for DNDP and FNA processes — the domain of the Energiforsk national methodology work (Source - Energiforsk 2026-1157 Nationell Metod Effekt och Kapacitetsprognoser (2026)).

Methods overview

Statistical methods

ARIMA/SARIMA (time-series; ARIMAX adds exogenous variables such as weather and time-of-day) are standard baselines for day-ahead STLF and MTLF. Multiple regression models handle stable medium-term patterns. All statistical methods have limited capacity for non-linear dynamics.

Machine learning

ANN, SVM/SVR, Random Forest, and XGBoost handle non-linearity and extract feature importance. More training data required than statistical methods. XGBoost is widely used in industry STLF due to speed and feature interpretability.

Deep learning

LSTM (Long Short-Term Memory) is the most studied DL method for STLF — its architecture captures long-range temporal dependencies naturally. CNN extracts local temporal patterns faster. BiLSTM captures bidirectional temporal context. Transformer models use attention mechanisms for non-sequential patterns.

Hybrid models — consistently best results

Combining architectures reliably outperforms any single model. CNN-LSTM, CNN-GRU, CNN-BiLSTM (Bayesian-optimized), and LSTM-Transformer are the most documented combinations. Quantified example: LSTM-BPNN achieved 1.47% MAPE vs 3.55% for standalone LSTM and 4.09% for standalone backpropagation network in a comparative study. (Source - Load Forecasting Methods Survey (2025))

Key input variables

Temperature is the dominant driver at all timescales — heating and cooling demands track outdoor temperature. Secondary inputs: historical load (capturing behavioral patterns), time features (hour-of-day, day-of-week, season, holidays), and economic activity indicators for MTLF/LTLF.

EV penetration and DR participation are increasingly important inputs as both reshape load patterns in ways that historical training data does not represent. An EV charging session adds 7–22 kW of local demand in patterns with no equivalent in pre-EV data.

Performance metrics

MAPE is the standard comparative metric across the literature. Hybrid models consistently achieve lower MAPE than single-method models across comparative studies.

Forecasting as a constraint on market design

The BRP imbalance channel

BRPs bear the financial cost of forecast error in the Nordic imbalance settlement framework. Poor STLF translates directly to imbalance charges. As the BSP/BRP framework extends to aggregated distributed resources (households, EVs, batteries), accuracy requirements cascade from large generators down to household resource pools.

The baseline problem

Explicit DR bids require a verified baseline — the counterfactual consumption absent flexibility activation. Baseline methodology is one of the most contested elements of Network Code on Demand Response design: an inaccurate baseline either inflates claimed DR delivery or fails to capture real curtailment. This makes load forecasting not just an operational input but an element of market integrity. See Baseline Methods for the full treatment.

Implicit DR synchronization at scale

When many BRPs respond to the same price signal with similar forecasting models, aggregate forecast errors can synchronize — amplifying system imbalances. This structural tension between implicit flexibility (the EU’s preferred pathway) and system operability is documented in Demand Response › Grid risks of demand response at scale.

Data challenges in flexibility contexts

Smart meter data quality: granular AMI data enables building- and device-level STLF, supporting real-time disaggregation and DR dispatch. But noise, missing values, and privacy constraints complicate model training. Low-voltage and microgrid forecasting has higher relative variability than bulk-system forecasting.

A missing-value gap is the visible failure mode; a flat estimated profile is the invisible one. Ei’s 2022–2023 supervision of ten DSOs’ metering compliance found 6 of 10 exceeded the legal time limit for missing values, and the two DSOs with long-running gaps attributed them mainly to smart-meter swaps that stalled mid-rollout: the old meter keeps measuring but the old communication infrastructure used to collect its data stops working as new meters are installed in the area, and meters delete stored values after a few months (the last ~3 months can typically be downloaded). Past that point the DSO must calculate values, usually distributing energy over periods using the customer’s earlier consumption profile — but Ei found some DSOs instead assign a single flat value across every quarter-hour, hour, or month for smaller customers. A stretch of “interval data” ingested into a forecasting model can therefore be synthetic and carry no real temporal information (a wiki inference; the mätföreskrifter require calculated values to be flagged as such, so whether the flag survives to a model’s training data is unverified) — a training-data contamination risk that a missing-value gap (which is at least visible as a gap) does not share. Separately, Ei received “strong indications” during 2023 that many DSOs report values at the wrong energy resolution — measuring at quarter-hour/hourly granularity internally but reporting monthly to the party downstream — meaning correctly measured data can still fail to reach a forecasting model at the resolution it needs. (Source - Ei Tillsyn Mätning och Rapportering (PM2024-02))

EV and DER data gaps: fleet composition, charging behavior, and V2G participation rates are inputs to EV load forecasting that are not yet available at the granularity needed for local grid-level planning.

Historical data invalidation: as load composition changes (EVs replacing ICE, heat pumps replacing gas boilers, large industrial electrification), historical patterns lose predictive value. Models trained on pre-2020 data will underperform for 2030 systems.

Swedish context

Sweden’s national DNDP methodology for load forecasting is developed by Energiforsk, with the most recent formalization in Source - Energiforsk 2026-1157 Nationell Metod Effekt och Kapacitetsprognoser (2026). This covers the LTLF/MTLF layer — capacity and peak demand projections used in DNDP and FNA. The STLF operational layer (day-ahead BRP and FSP market participation) is less standardized and remains a commercial and technical competency developed by individual aggregators, BRPs, and energy service providers.

The DHV data architecture and FIS infrastructure under Network Code on Demand Response will eventually provide the standardized metering and settlement data that makes automated STLF-based DR participation scalable. Until then, the data access challenge remains a structural barrier to household-level STLF.

AI-driven load forecasting — sector capability gap

The PREDATOR project (Energiforsk 2026:1168, Source - Energiforsk 2026-1168 AI-modeller Prognostisering Efterfrågan El (2026)) conducted four workshops with eight Swedish DSOs (Ellevio, Göteborg Energi, Mölndal Energi, Trollhättan Energi, Jönköping Energi, Umeå Energi, C+ Energi, Karlstads Kommun) between November 2023 and April 2024 to understand what data-driven demand forecasting tools they need and what they currently have.

Key finding: Many Swedish DSOs have no internal data-driven demand or load forecasting models at all — the report attributes this to many DSOs lacking the established infrastructure to integrate predictions of high-load technologies (EVs, heat pumps) into planning, not just to a shortage of modeling sophistication. Even DSOs with some quantitative forecasting lack integrated end-to-end solutions that combine DER inventory identification, behavioral profiles, and growth scenario modeling.

What DSOs actually want: DSOs do not want partial tools or specialized components — they want complete, integrated end-to-end solutions that handle the full forecasting workflow.

“Analyze-aggregate-extrapolate” framework: RISE Research Institutes of Sweden proposed a conceptual architecture for holistic DSO demand forecasting:

1. Analyze — identify and inventory current high-load technologies (EVs, solar, heat pumps, price-control apps like Tibber) by city/quarter
2. Aggregate — build current-state load profiles per city/quarter from that inventory
3. Extrapolate — apply the existing RISE prediction model to project adoption up to 10 years ahead; Sub-report 2 plans to broaden the model’s data with survey/statistics sources (e.g. SCB, Eurobarometer, ESS — 17 candidates considered) and economic indicators (household income, inflation, interest rates, etc.)

Follow-up: a larger project DESGRID is planned to implement and validate this framework. A Vinnova Advanced Digitalization funding application was postponed rather than submitted — the call’s 50% in-kind co-funding requirement meant several interested DSOs couldn’t confirm commitments in the short window before Swedish summer holidays; a better-timed retry is planned for 2025.

This gap is directly documented by Energiforsk 2026:1157 (Source - Energiforsk 2026-1157 Nationell Metod Effekt och Kapacitetsprognoser (2026)) from a different angle: survey of 50 DSOs found 28% lack documented methodology, 40% lack documented power templates, and 36% do not validate systematically — with several noting their forecasts have systematically overestimated actual power needs.

A named, large-DSO case confirms the pattern rather than being an exception to it. A 2023 Uppsala University thesis, written from inside Vattenfall Eldistribution with a company supervisor and mentor, evaluated the load forecasts used in CoordiNet‘s Uppsala flexibility market and found that CoordiNet was the first time Vattenfall Eldistribution had created load forecasts for its grid. Developing and running the forecasts was outsourced to Expektra, and — stated as the thesis’s own starting problem, not an external criticism — that outsourcing left the company without a detailed internal understanding of how its own forecasts worked, including basic operational facts like how the day-ahead and intraday models actually differ. The thesis’s first phase was a round of interviews and an interactive presentation to map that gap, before any error analysis. Vattenfall is one of Sweden’s three largest DSOs; the capability gap the PREDATOR and Energiforsk 2026:1157 surveys document at a sector level is not confined to small or mid-size companies. (Source - Wiss Utvärdering av Lastprognoser CoordiNet Uppsala (2023))

Build vs. buy — the Swedish forecasting vendor landscape

A DSO without internal STLF capability has three paths, all in active use: buy commercial software, buy a forecasting service outright, or adopt an open-source toolkit.

• Commercial software: Vitec Energy’s Aiolos Forecast Studio covers intraday through multi-year horizons for electricity, gas, and district heating/cooling, and is under active development (a modern API and battery-optimization functionality as of its most recent release).
• Forecasting as a service: Expektra is the concrete example above — per the Wiss thesis, it built and ran Vattenfall Eldistribution’s CoordiNet forecasts rather than the DSO building them itself. (Expektra’s own current website markets an AI-driven power-trading platform and BRP-as-a-service, not a distinct grid-company forecasting product.)
• Open source: OpenSTEF, hosted at LF Energy, originated at Dutch DSO Alliander specifically to combat distribution-grid congestion and was open-sourced to become a shared industry alternative to proprietary tools — a real option for a DSO wary of vendor lock-in but not ready to build a full STLF stack from scratch.

None of these paths resolves the understanding gap on its own — the Wiss thesis suggests that buying the forecast without also building internal literacy about how it works just relocates the capability gap rather than closing it.

EV fast-charging station peak-load forecasting without history (Lund/Öresundskraft)

A 2026 Lund University study, using Öresundskraft’s own fast-charging station data (33 DC stations in northwestern Skåne, Jan 2025–Mar 2026), addresses the specific cold-start version of the “EV and DER data gaps” problem above: predicting a planned charging station’s 24-hour peak-load profile from pre-construction metadata alone (capacity, charging-point count, location, proximity to highway) — before the station has generated a single hour of its own load data. The method splits load-shape prediction (K-means clustering + random-forest classification) from peak-magnitude prediction (quantile random forest), producing a risk-adjustable profile: planners choose how conservative a capacity allocation to make, trading overload risk against unused capacity — a newsvendor-style framing of the same connection-planning decision Grid Capacity Utilization already covers from the DSO side. The conservative (99th-percentile) prediction covered observed peaks for 93.4% of stations without defaulting to theoretical-maximum overkill; smaller stations (fewer charging points) proved harder to predict, since a handful of simultaneous charging sessions can push them close to full capacity, while larger stations’ peaks are naturally damped by the low probability of many vehicles charging at once. (Source - Gerholm Lindstrom EV Fast-Charging Peak-Load Forecasting (2026))

Data gaps

• DSO-side STLF practice in Swedish LFMs — CoordiNet’s answer is now known (Vattenfall Eldistribution fully outsourced to Expektra; see Build vs. buy above); sthlmflex, SWITCH, and Effekthandel Väst DSOs’ own practice remains open — what models, what input data, what forecast horizon
• FSP/aggregator STLF practice — what methods are deployed in practice for SWITCH, NODES, and Effekthandel Väst market participation

Related pages

• Demand Response — what STLF enables: explicit DR participation, FSP bidding, aggregator dispatch
• Balancing Markets — BRP planning and imbalance settlement; reserve procurement that DR participates in
• Flexibility Market — FSP/BRP bidding; baseline methodology as product design challenge
• Baseline Methods — dedicated page on baseline measurement and verification
• Distribution Network Development Plan — DNDP load scenarios built on LTLF/MTLF
• Flexibility Need Assessment — FNA flexibility needs quantification uses LTLF scenarios
• Network Code on Demand Response — baseline methodology regulation; SP qualification requirements
• Distribution System Operator — the DSO capacity-planning use of LTLF/MTLF forecasting this page’s AI-adoption section examines