Digitala Monster — AI och Svenska Energisystemet (2026)

“Digitala monster — En analys av AI:s effekter på det svenska energisystemet.” Kasper Persson, master’s thesis, Miljö- och energisystem, Lunds Tekniska Högskola, June 2026 (ISRN LUTFD2/TFEM—27/5251—SE). Conducted in collaboration with Energiforsk.

Method: literature review plus 17 interviews with 18 people, of which only 14 were with the external Swedish energy-sector actors this page describes (DSOs, Svenska kraftnät, RISE, industry vendors) — the remaining 3 were scoping interviews with Energiforsk’s own programme/council staff, used to identify which external actors to approach rather than as technical sources in their own right. Focused on three areas identified as where AI will have the greatest impact: forecasting, flexibility services, and digital twins, plus a dedicated risk chapter. Deliberately Sweden-focused, on the grounds that Sweden’s district-heating-heavy, high-renewables system is under-represented in international AI/energy literature.

Caveat on sample bias (ch. 8.3, Felkällor): the author explicitly flags that the interview sample is small and self-selected, drawn almost entirely from Energiforsk’s own contact network, likely skewed toward technologically mature and AI-optimistic actors, and likely under-representing AI-critical perspectives. He also notes conclusions may already be dated by publication given how fast AI development moves. This caveat applies to interview-derived claims throughout this summary (e.g. “spot price as the strongest predictor,” VPP/digital-twin assessments) — treat them as one practitioner sample’s view, not a representative survey.

Forecasting (ch. 4)

Short-horizon consumption forecasting: one interviewed large grid owner (Intervju 8) has substantially improved system-level consumption forecasts using ML on solar irradiance, spot prices, temperature and historical use — spot price is identified as the single strongest predictor, especially in areas with heavy industry, most visibly during industrial downturns. This gives DSOs better grounds for deciding which production resources to call or curtail; a second interviewee (Intervju 12) confirms smaller DSOs are also getting more proactive, concrete guidance on which reserve power to activate and when.

Long-horizon consumption forecasting: covered by Source - Energiforsk 2026-1168 AI-modeller Prognostisering Efterfrågan El (2026) — the DESGRID/PREDATOR scoping project. A difficulty flagged in both sources: household technology-adoption decisions (EVs, heat pumps) are influenced by political decisions (subsidies), which can date historical training data when a subsidy scheme changes.

Renewables production forecasting: described as a “two-stage rocket” — AI first improves the underlying weather forecast (hybrid physical/ML models), then a second forecasting layer converts that into a production forecast. Strengthening either stage improves grid-frequency management by extension.

Price forecasting: as of 2024 still primarily traditional statistical methods, not ML; the author found no Swedish interviewee positioned to speak to AI’s current role here specifically — flagged as an open question.

Market-optimisation example: one interviewed company (Intervju 17) uses AI to run large numbers of forecasts across multiple ancillary-service markets and has automated its battery-park buy/sell decisions — batteries charge when price is low/supply abundant and discharge when price is high/supply short, which the company frames as a direct grid-stabilising side effect of its trading strategy.

Imbalance forecasting and the aFRR/mFRR activation-order reform: Svenska kraftnät continuously produces imbalance forecasts using a mix of predictive AI, statistics and physical models; for these, too much historical data was found to be counterproductive since grid conditions change too fast for old data to stay relevant.

Since 2025, Sweden’s aFRR/mFRR activation order has reversed. Historically aFRR activated reactively first to restore frequency, with mFRR called in afterward to relieve it. Now mFRR activates first and proactively: Svenska kraftnät produces 5-minute-resolution frequency-error predictions (the thesis attributes this cadence to Svk specifically, then separately notes competing forecasting tools — both AI-based and traditional — are continuously scored against realised outcomes), and the best-performing forecast determines how much mFRR to activate on a 15-minute bid cycle; aFRR then only regulates the residual. Reasons given: (1) alignment with the EU-mandated cross-border platforms Sweden joins in 2027 (MARI for mFRR, PICASSO for aFRR), and (2) mFRR bids are cheaper than aFRR bids, so proactive mFRR use lowers activation cost. This is an operational detail not previously in this wiki’s Balancing Markets coverage.

Predictive maintenance: ML on IoT sensor data (both production units and grid infrastructure) shifts maintenance from age-based/reactive to condition-based, catching component degradation before failure — cited as valuable partly because newer equipment often has shorter service life than older generations, undermining age-based scheduling logic.

Author’s assessment: forecasting is the domain where AI has had the largest measured impact in the Swedish energy system so far. The gain is not “AI magic” but sheer data-handling capacity — more frequent, more data-rich forecasts, shifting parts of the system from reactive to proactive operation. The binding constraint is data access and quality, not algorithms; time-series data has a practical “best-before date” given how fast the system itself is changing, and confidentiality restricts sharing between actors.

Recommendations to Energiforsk (not adopted as wiki fact, noted as proposals): (1) federated learning for shared forecasting models across competing market actors without pooling raw training data — used broadly elsewhere, unexplored in the Swedish energy sector; (2) Kaggle-style public forecasting competitions, with Energiforsk organising and supplying data rather than running a full platform itself.

Flexibility services (ch. 5)

Smart control of behaviour-driven load (space heating, hot water) is already AI-tractable because the underlying human-activity patterns are strongly regular; this in turn frees capacity to shift non-behaviour-driven load (fridges, ventilation, heat pumps run in the background) without noticeably affecting comfort — e.g. boosting a heat pump early morning, then cutting it before an evening price/load peak.

V2G: by 2050 Sweden’s vehicle fleet is expected to be almost entirely electrified. Smart charging (e.g. charging overnight when load is low) is already common; discharging back to the grid is technically mature but has only reached pilot scale — see Vehicle-to-Grid and Source - Så Kan V2G Gå Från Projektform Till Marknaden (2025), cited directly in this thesis. AI’s role is making smarter charge/discharge timing decisions from more variables (weather, price) — though the author is explicit that AI is not necessarily the only or best method here.

VPP aggregation: individually negligible resources (one EV, one solar roof) become a grid problem in aggregate if thousands act simultaneously; virtual power plants coordinate them so their individual strengths/weaknesses offset each other. VPPs are framed as requiring advanced AI by construction, since they combine forecasting, scheduling, production optimisation and multi-market bidding on large data volumes.

Sector coupling — district heating as an ancillary-service resource: district heating (CHP plants, large heat pumps, electric boilers, accumulator tanks acting as thermal storage) is a plausible flexibility resource, but the thesis found Swedish district-heating firms currently active only on day-ahead/intraday markets, with no evidence of ancillary-service participation and no identified AI use for that purpose — flagged as a real gap given rising district-heating prices and competitiveness concerns raised independently by interviewees.

Author’s assessment: flexibility services is a fast-growing market pulling in new AI-native entrants, but technology maturity varies enormously by application and cannot be generalised. Recommendations to Energiforsk: quantify the AI/digitalisation savings case for district heating networks specifically (general grid literature already shows 27–80% investment-cost savings from integrating flexibility instead of building out infrastructure — Smart Energy Europe & DNV, 2022 — but this hasn’t been demonstrated for district heating); demonstration/scenario projects on digitalisation vs. hardware build-out; compensation/auction-model concepts for V2G and large-building flexibility (with an explicit caveat to first check whether AI is even the right tool versus traditional optimisation methods); and Swedish VPP landscape-building, since the concept is discussed internationally (Germany’s Statkraft VPP cited as a working example) but only lightly present in Sweden (Greenely, Tibber use the term but at limited scale).

Digital twins (ch. 6)

Definition used, quoting the thesis’s own Swedish rendering of ENTSO-E’s definition (this document gives no English original, so the English here is a back-translation, not a verified verbatim ENTSO-E quote): a dynamic, virtual counterpart of an object, system or process, continuously updated with data from its real-world counterpart (ENTSO-E, 2026).

Most existing Swedish digital twins are small, single-purpose commercial products — e.g. one interviewed company’s digital twin of relay protection within a customer’s grid area, and a separate district-heating company’s digital twin of building heat demand for property owners.

Svenska kraftnät’s Kraftsystemhubben is presented as the leading Swedish example, functioning as a “system of systems”: it ingests data from DSOs and foreign TSOs and routes it to whichever internal Svk system needs it — including feeding the imbalance-forecasting algorithm behind the aFRR/mFRR reform above. Whether it strictly qualifies as a “digital twin” is an open internal debate the thesis declines to resolve, noting only that it has digital-twin-like functionality either way. Svenska kraftnät‘s wiki page currently records only its budget line (11 MSEK, closing mid-2027); this source adds the functional description.

RISE national grid digital twin: funding approved 30 April 2026 (largely funded by Vinnova, with Energiforsk and Svenska kraftnät as partners — the thesis doesn’t state Vinnova leads/directs the project, only that it provides most of the funding) for RISE to build a national-scale grid digital twin. Early stage at time of writing; framed as a research “laboratory” for testing different starting conditions and scenarios, not an operational tool.

Success factors identified by interviewees: (1) design around one clearly defined use case — projects framed as “build a digital twin” in the abstract consistently fail; (2) build bottom-up from the smallest relevant components rather than attempting an all-encompassing model, while still working toward common data-handling standards early so separately-built twins can eventually interconnect.

Author’s assessment: digital twins are an early, hype-adjacent concept in Sweden; today’s examples are small and specialised, and it’s genuinely uncertain whether the large interconnected systems described in international literature will materialise. Cost of the underlying IoT sensor infrastructure required for real-time updating is flagged as an often-overlooked line item that could itself become a barrier. Recommendations to Energiforsk: help define good use cases; convene a cross-industry digital-twin working group (potentially split by electricity vs. heat systems) partly to seed early data-standard agreement; commission an economic analysis of the sensor/IoT cost base.

Risks (ch. 7)

Competence and knowledge: a widening gap between available and industry-demanded technical skill (echoing TechSverige, 2026), compounded by (a) organisational over-trust in AI as inherently superior without understanding its data-quality dependence, (b) a generation of power/electrical engineers lacking IT/AI grounding being handed fast-rolled-out tools, (c) cultural resistance to changing established ways of working in a historically conservative, asset-management-focused sector, and (d) a risk that AI automates away the “simple” tasks that historically trained junior staff to catch errors — leaving nobody positioned to notice when the AI itself is wrong. This burden falls hardest on small and mid-sized companies, who lack the resources to compete for talent or train internally, risking competence concentration at large incumbents.

Explainability: a recurring theme, especially among more IT-literate interviewees. Two stated reasons it matters beyond technical hygiene: accountability (a “black box” wrong decision cannot be attributed or defended), and detectability (an unexplainable model makes it harder to notice if it has been compromised). The interviewed AI-driven trading company explicitly insists a human must always be able to review and understand its trading decisions.

Economic risk: (1) not starting digitalisation in time — smaller firms lack the competence to begin, face a lag before data-driven digitalisation pays off, and face rising component costs the longer they wait; (2) investing in the wrong infrastructure — poor procurement competence can lead to over/under/irrelevantly-specified sensor deployments wasting capital.

Digital security: beyond conventional cybersecurity, the thesis names data poisoning as an AI-specific risk — corrupting training data or manipulating live inputs to induce hallucination/incorrect output, either during training or in production. Currently AI mostly informs decisions rather than directly controlling systems in the Swedish grid, but the risk rises as more directly-actuating AI software is connected. A second privacy-adjacent risk: the fine-grained consumption data high-resolution smart meters and IoT devices generate could, if leaked, expose behavioural patterns of individuals or firms.

Dependency on foreign suppliers — two distinct mechanisms: (1) cloud infrastructure hosted on servers in a state that could become adversarial, which could become unavailable in a conflict; (2) hardware from a foreign manufacturer capable of pushing firmware updates post-installation — named risk scenario: malicious firmware pushed to widely-deployed hardware specifically named as V2G-connected EVs and grid-tied inverters, in a state-conflict scenario.

Political barriers: regulation lags the technology and still rewards classical grid build-out over digital/flexibility solutions that would often be more cost-effective — a barrier interviewees raised repeatedly.

Author’s assessment: most identified risks trace back to the pace of AI development itself — actors making unsound investments out of fear of falling behind, whether through misapplied AI or genuinely unexplainable deployments; regulation struggling to keep pace compounds this. The single largest identified barrier to AI adoption is the competence question (ch. 8.2), not technology maturity or economics.

Discussion and conclusions (ch. 8–9)

AI is framed as a precondition for successful electrification, not merely a nice-to-have — the emerging grid will be too information-dense for human operators to monitor unaided, but tools must be held to high standards of transparency and reliability rather than deployed reflexively. The AI “bubble” discussion in contemporary media is characterised as being about generative AI specifically, distinct from and not casting doubt on the forecasting/optimisation/control tools discussed in this thesis, which the author expects to persist regardless. Competence is named as the decisive bottleneck; data access/quality as the second. The thesis explicitly used generative AI (Microsoft Copilot, ChatGPT) only for reference-list formatting, pre-screening sources for relevance before full reading, getting tips on where to search for relevant academic sources, and light language feedback — never for original text.

Relevance to this wiki

• Adds a dateable, specific operational reform to Balancing Markets: the aFRR/mFRR proactive-mFRR-first activation order since 2025.
• Substantially deepens Svenska kraftnät‘s existing (budget-line-only) mention of Kraftsystemhubben with functional detail.
• Gives Digitalization and Smart Grid concrete Swedish digital-twin examples (the RISE national project, Kraftsystemhubben’s ambiguous digital-twin status) and a risk register (competence, explainability, market-power centralisation, foreign-hardware dependency) that page currently lacks.
• Supplies a named, attributable DSO forecasting insight (spot price as the strongest consumption-forecast predictor) not previously sourced anywhere in the wiki.
• Corroborates and cross-references both Source - Energiforsk 2026-1168 AI-modeller Prognostisering Efterfrågan El (2026) and Source - Så Kan V2G Gå Från Projektform Till Marknaden (2025), which the thesis itself cites.
• Flags district-heating ancillary-service participation as an identified, sourced gap — not previously tracked as a data gap anywhere in this wiki.