Residential demand response: Experimental evaluation and comparison of self-organizing techniques

Residential demand response (DR) has gained a significant increase in interest from industrial and academic communities as a means to contribute to more efficient operation of smart grids, with numerous techniques proposed to implement residential DR programmes. However, the proposed techniques have been evaluated in scenarios addressing different types of electrical devices with different energy requirements, on different scales, and have compared technique performance to different baselines. Furthermore, numerous review papers have been published comparing various characteristics of DR systems, but without comparing their performance. No existing work provides an experimental evaluation of residential DR techniques in a common scenario, side-by-side comparison of their properties and requirements derived from their behaviour in such a scenario and analysis of their suitability to various domain requirements. To address this gap, in this paper we present four self-organizing intelligent algorithms for residential DR, which we evaluate both quantitatively and qualitatively in a number of common residential DR scenarios, providing a performance comparison as well as a benchmark for further investigations of DR algorithms. The approaches implemented are: set-point, reinforcement learning, evolutionary computation, and Monte Carlo tree search. We compare the performance of approaches with regards to energy-use patterns (such as reduction in peak-time energy use), adaptivity to changes in the environment and device behaviour, communication requirements, computational complexity, scalability, and flexibility with respect to type of electric load to which it can be applied, and provide guidelines on their suitability based on specific DR requirements.