Demand Side Management in the Smart Grid

Abstract

The global effort towards a decarbonised energy sector led to an increased share of renewable energies in the energy mix of most industrialised countries. Production from renewable resources as wind and solar is intrinsically intermittent, and often installed at medium and low voltage grid. This, together with the growing urbanisation, the in- creasing popularity of electricity-based climate control systems and electricity-based private transportation, challenges the distribution systems operators to run the low voltage grids with small safety margins. Depending on the customer and the specific application, a portion of the electricity demand (or production) is flexible, and can eventually be used to support the system stability. Such practice is called Demand-side Management (DSM), and flexible units are called Distributed Energy Resources (DERs), if they only produce energy, or Demand Side Resources (DSRs), in the case they can also (or only) consume. This research investigates a specific control approach to DSM, called Direct Load Control (DLC). DLC is based on the assumption that the controlled units react in a predictable way to received control signals, and notify a failure if their operation is compromised. In this research DLC is investigated at unit level and at aggregation level by means of model-based predictive control and model-free online control. Predictive control allows maintaining the flexibility of DSRs longer than online control, and allows pursuing quality of service (QoS) objectives towards both customers and SOs. However, it needs a model of the controlled units to optimise their operation over time, and field data to monitor the units’ operation. On the other hand, the online control here investigated takes myopic decision based on field data, without the need of unit model. Therefore it does not guarantee optimality in the long run. Two approaches are investigated within predictive control: the first is based on grey-box modelling and quadratic Model Predictive Control (MPC), while the second is based on Artificial Intelligence modelling and gradient-free optimisation. This research contributes to the first approach with a novel coordination scheme for a cluster of DSRs, which innovates with respect to other solutions proposed in literature by allowing independent design of the units local controllers and the use of heterogeneous control signals, such as continuous, integer or binary. The second approach foresees the application of self-learning models to predictive control. In this context, the contribution of this research is twofold: first it proposes the application of ensemble methods and decision trees to modelling of climate control systems, then it provides experimental validation of a combined Reinforcement Learning/in-domain knowledge approach for DSM application. The second approach to DSM in based on model-free online control. The proposed controller uses only field measurements and a generic domain knowledge to abstract the unit flexibility for DSM applications. It allows real-time control of power flow in single DSRs and, since there is no operation planning involved, decisions are taken online on the basis of field measurements. In this context, a novel approach is proposed to control thermostatically-controlled loads (TCLs) within a framework for control of micro grids by means of explicit power set points.