Accurate and robust forecasting is a critical challenge for the reliable operation and optimization of smart microgrids, especially as they increasingly integrate renewable energy sources and rely on edge computing infrastructure. Uncertainty from weather, rapidly changing consumption patterns, and the resource limitations of edge devices hinder the deployment of conventional forecasting models, which typically assume stationary data and require large, clean datasets. This licentiate thesis addresses these gaps by developing a comprehensive and adaptive pipeline for time-series forecasting in microgrids, with a focus on resilience to uncertainty, concept drift, data scarcity, and resource constraints. The research presents a suite of novel and scalable approaches, including clustering-based pre-processing, adaptive deep learning models, probabilistic and uncertainty-quantifying forecasting methods, federated learning, and instance-based transfer learning. These techniques are designed to deliver high predictive accuracy and adaptability under non-stationary conditions, while remaining lightweight enough for deployment on constrained edge devices. Privacy-preserving and collaborative mechanisms, such as federated and transfer learning, enable model generalization and robustness across heterogeneous, distributed microgrid environments. Building on seven peer-reviewed publications, this work establishes a unified pipeline for intelligent forecasting at the grid edge. Key contributions include: (1) dynamic drift-adaptive learning frameworks for real-time adaptation, (2) probabilistic and uncertainty-quantifying forecasting for risk-aware energy management, (3) multi-step prediction strategies for short- and medium-term forecasting, and (4) clustering-guided transfer learning to enhance prediction in data-scarce scenarios. Validation on both real-world and simulated microgrid datasets demonstrates substantial gains: for example, the proposed transfer learning approach improves short-term forecasting accuracy in data-limited buildings by up to 40%, while adaptive and federated models achieve robust performance with reduced computational and communication overhead compared to centralized baselines. Overall, this thesis advances the state-of-the-art in interpretable, scalable, and resilient forecasting for smart microgrids, supporting the transition to adaptive, decentralized, and privacy-aware energy systems capable of thriving in dynamic and uncertain environments.

Noggrann och robust prognostisering är en avgörande utmaning för tillförlitlig drift och optimering av smarta mikronät, särskilt i takt med att dessa i allt högre grad integrerar förnybara energikällor och förlitar sig på edge computing-infrastruktur. Osäkerheter till följd av väder, snabbt föränderliga konsumtionsmönster och begränsade resurser hos edge-enheter försvårar implementeringen av konventionella prognosmodeller, som vanligtvis antar stationära data och kräver stora, rena datamängder. Denna licentiatuppsats adresserar dessa utmaningar genom att utveckla en omfattande och adaptiv pipeline för tidsserieprognostisering i mikronät, med fokus på robusthet mot osäkerhet, konceptdrift, databegränsning och resursknapphet. Forskningen presenterar en uppsättning nya och skalbara metoder, inklusive klustringsbaserad förbehandling, adaptiva djupinlärningsmodeller, probabilistiska och osäkerhetskvantifierande prognosmetoder, federerad inlärning samt instansbaserad transferinlärning. Dessa tekniker är utformade för att uppnå hög prediktiv noggrannhet och anpassningsförmåga under icke-stationära förhållanden, samtidigt som de är tillräckligt resurseffektiva för att köras på begränsade edge-enheter. Sekretessbevarande och samarbetsinriktade mekanismer, såsom federerad och transferinlärning, möjliggör modellgeneralisering och robusthet över heterogena, distribuerade mikronätsmiljöer. Baserat på sju sakkunniggranskade publikationer etablerar detta arbete en enhetlig pipeline för intelligent prognostisering vid elnätets ytterkanter. Viktiga bidrag inkluderar: (1) dynamiska, driftanpassade inlärningsramverk för realtidsanpassning, (2) probabilistisk och osäkerhetskvantifierande prognostisering för riskmedveten energihantering, (3) flerstegsprognoser för kort- och medellång sikt, samt (4) klustringsstyrd transferinlärning för att förbättra prognoser i dataskrala scenarier. Validering på både verkliga och simulerade mikronätsdatamängder visar på betydande förbättringar: till exempel förbättrar den föreslagna transferinlärningsmetoden korttidsprognosens noggrannhet i databegränsade byggnader med upp till 40%, medan adaptiva och federerade modeller uppnår robust prestanda med lägre beräknings- och kommunikationsbörda jämfört med centraliserade baslinjer. Sammanfattningsvis för denna uppsats forskningsläget framåt inom tolkbar, skalbar och resilient prognostisering för smarta mikronät, och stödjer övergången till adaptiva, decentraliserade och integritetsmedvetna energisystem som kan frodas i dynamiska och osäkra miljöer.

Accurate forecasting is vital for reliable operation of smart microgrids, especially with growing renewable integration and edge computing reliance. Conventional models falter under uncertainty from weather, shifting demand, and limited edge resources, as they assume stationary data and large clean datasets. This licentiate thesis develops an adaptive forecasting pipeline tackling uncertainty, concept drift, data scarcity, and constraints. It introduces clustering-based preprocessing, adaptive deep learning, probabilistic forecasting, federated learning, and transfer learning. These methods deliver accuracy and adaptability under non-stationary conditions while remaining lightweight for edge deployment. Privacy-preserving mechanisms enhance robustness across distributed microgrids. Based on seven peer-reviewed publications, contributions include: (1) drift-adaptive learning for real-time updates, (2) probabilistic forecasting for risk-aware management, (3) multi-step prediction for short- and medium-term horizons, and (4) clustering-guided transfer learning for scarce data. Validation on real and simulated datasets shows major gains: transfer learning improves short-term accuracy by up to 40%, while adaptive and federated models sustain robust performance with reduced overhead.

The growing scale and heterogeneity of distributed energy systems has increased the complexity of their operational energy management. These systems must operate reliably under limited data and exogenous uncertainty while satisfying technical, social, and economic constraints. Classical optimization provides strong theoretical guarantees but often relies on deterministic models and point forecasts. Conversely, learning-based methods offer adaptability and scalability but may lack explicit safety handling and explainability, which are critical for deployment in safety- and flexibility-critical energy infrastructures.

Consequently, this thesis establishes an uncertainty-aware Energy Management System (EMS) architecture that preserves feasibility across planning and execution while enabling safety-constrained and human-centric intelligent control. Within a unified layered EMS structure, we address (i) exogenous uncertainty in generation and load, (ii) epistemic uncertainty arising from limited data, and (iii) exploration-related uncertainty during policy learning while supporting prosumer-in-the-loop operation. We first develop robust optimization methods with flexible and asymmetric uncertainty sets to obtain less conservative schedules while maintaining feasibility guarantees and improving economic performance. We further establish a safety-constrained reinforcement learning framework and benchmarking environment that enable systematic evaluation of performance–safety trade-offs. Finally, building on the EMS planning layer, we introduce an explainability and interaction layer in which schedules are complemented with solver-grounded explanations and what-if analyses, supporting human-in-the-loop flexibility decisions.

By jointly advancing robustness, data-efficient forecasting, safety-constrained learning, and explainable decision support, this thesis contributes an integrated and trustworthy decision-making paradigm for decentralized energy systems.

As decentralized energy resources proliferate, energy management system (EMS) design needs to handle stochastic uncertainty, hard operational constraints, and stakeholder-oriented decision support. These requirements are often addressed in isolation, resulting in a lack of integrated approaches that jointly provide constraint satisfaction, scalable operation, and transparent decision rationales.This thesis advances a unified EMS design perspective spanning three coupled dimensions: (i) robustness to multi-dimensional uncertainty, (ii) constraint-aware adaptive control, and (iii) explainable, deployable decision support for prosumer engagement. Specifically, the thesis develops an end-to-end EMS stack combining edge–cloud coordination, uncertainty-aware scheduling, safe reinforcement learning, and generative-AI explanations grounded in the underlying optimization model. Validated across microgrid, building, and energy-community case studies, the thesis establishes an empirically grounded pathway toward trustworthy, intelligent energy systems capable of accelerating the energy transition.