Multi-Layered Modelling Knowledge Graph and Response Derivative Framework for Autonomous Buildings, Energy Communities, and Positive Energy Districts

This record provides the Multi-Layered Modelling Knowledge Graph (MLM KG), a machine-readable, semantically structured knowledge artefact that formally encodes the MLM framework, the Response Derivative (RD) concept, and associated regulatory compliance mappings developed by the Energy Informatics Group (EIG) at Tyndall National Institute, University College Cork. The MLM KG is the canonical implementation reference for the multi-layer modelling framework introduced in: O'Regan B., Tahir F., Mould K., O'Leidhin E. (2026). Towards Autonomous Buildings, Communities and Positive Energy Districts: Multi-Layer Modeling and Edge-Enabled Islanding for the Energy Transition. International Energy and Environment Building Science Conference (IEECB&SC'26). O'Regan B. (2025). From Flexibility to Trading: Optimizing Electricity & Heat in O-CEI Pilot 1 — Demonstrating Symbiotic Buildings and Market-Ready Flexibility Services. IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT-Europe 2025), Valletta, Malta, October 20–23, 2025. Scientific Context Modern electricity grids face increasing volatility from renewable intermittency, climate-driven disruptions, and the electrification of heat and transport. The Aran Islands, Ireland, a pilot site of the O-CEI (Open Cloud-Edge-IoT) project, experienced extended power outages during Storm Éowyn (January 2025, 184 km/h gusts at Mace Head), a real-world demonstration of the vulnerability of isolated communities to grid disruption and the urgent need for resilient, locally autonomous energy systems. The MLM framework addresses this by enabling buildings and communities to shift from passive consumers to active, intelligent energy agents, capable of self-optimisation, peer-to-peer energy trading, and autonomous islanded operation during grid disturbances. The Multi-Layer Modelling Framework The MLM framework integrates four complementary computational layers into a unified hierarchical architecture: Deterministic Layer enforces physical and operational feasibility through thermodynamic models (RC thermal networks, NTU-ε heat exchangers, COP curves), electrical constraints (Kirchhoff's laws, voltage and frequency limits), comfort constraints (temperature 19–24°C, CO₂ < 1000 ppm), and safety limits (battery SoC 20–95%). Stochastic Layer quantifies uncertainty from renewable variability, occupant behaviour, and market volatility using weather ensembles, Markov chain occupancy models, ARIMA price forecasting, and Monte Carlo simulation. This layer generates the probability distributions and scenario sets required for robust decision-making. AI/ML Layer enhances predictive performance and computational efficiency through LSTM networks for short-term load and generation forecasting, neural network surrogate models replacing computationally expensive physical simulations at 100–1000× speedup, clustering algorithms for pattern identification, and federated learning for privacy-preserving cross-site model improvement. Reinforcement Learning Layer introduces adaptivity through Q-learning, DQN, and Actor-Critic methods. The RL agent optimises control policies, load shifting, storage dispatch, heat pump scheduling, islanding decisions, within the feasibility bounds established by the deterministic layer, informed by stochastic uncertainty quantification and AI-enhanced predictions. These four layers are deployed across a distributed TinyML–edge–cloud architecture: far-edge building devices (Jetson Nano running PARA//EL) for ultra-low latency deterministic control; community edge nodes (running EdgeWare) for 15-minute optimisation cycles and P2P trading settlement; and the FLEXUS cloud platform for regional coordination, federated learning, and market integration. Response Derivative: Novel Theoretical Contribution A key original contribution of this work is the Response Derivative (RD), introduced by Brian O'Regan as a novel metric for quantifying system responsiveness in distributed, probabilistic energy systems. Unlike traditional capacity-based metrics (nameplate power ratings, installed capacity), RD measures how quickly and reliably a system can respond to control signals under uncertainty. Formally: RD = δ(actual response) / δ(stimulus) | context RD is estimated within the stochastic and AI layers, bounded by deterministic constraints, and consumed by the RL layer as a primary decision primitive. For a given asset or portfolio, RD captures the sensitivity of system response to control inputs, accounting for: Physical response characteristics: thermal inertia, ramp rate limits, efficiency curves Behavioural patterns: historical occupant override frequency, comfort preference variability Market conditions: price-responsiveness of flexible loads, opportunity costs Uncertainty bounds: confidence intervals around expected responses The RL governance mechanism uses RD to evaluate and derate candidate actions based on expected responsiveness. Actions with low RD (unreliable or negligible response expected) are avoided or scaled down, preventing the system from relying on flexibility that will not materialise. This ensures autonomous decisions remain safe and effective as system conditions evolve. In the LP/MILP implementation, RD is computed as the shadow price (dual variable) of the optimisation constraints, the marginal value of relaxing each constraint, providing a mathematically rigorous operationalisation of the concept. Supporting Infrastructure minutemanSEM provides continuous minute-level semantic health and service readiness assessment using structured S:E:M encoding (Status:Equipment:Message) across five ESRF levels from ESRF-5 Optimal (≥99.0% health) to ESRF-1 Critical (