Optimization heuristics for residential energy load management

The MS thesis is concerned with the problem of scheduling the daily energy loads in a multihouse environment from the point of view of an energy retailer. We assume that the residential users own a set of home appliances (washing machines, dishwashers, ovens, microwave ovens, vacuum cleaners, boilers, fridges, water purifiers, irons, TVs, personal computers and lights) that are supposed to be used during the day. Houses can also be equipped with Photo Voltaic (PV) panels, which produce energy in a discontinuous way, and batteries that allow the system to store and release energy when required. The day is subdivided into 96 timeslots of 15 minutes each. For each appliance, we suppose to know the load profile, that is, a set of successive timeslots with the corresponding amount of energy required. Given the load profile of each appliance, the time windows in which the appliances must be executed, the physical characteristics of the batteries, the energy amount produced by the PV systems, the problem is that of scheduling the various appliances (assigning their starting timeslots) so as to minimize an appropriate objective function while respecting the maximum capacity of the meters (usually 3 kW). We consider minimizing the total maximum peak. This Residential Energy Load Management Problem is a challenging extension of the classical Generalized Assignment Problem (GAP). Since the Mixed Integer Linear Programming (MILP) formulation can be solved within reasonable computing time only for small instances, we developed various methods to tackle medium-to-large size instances: a Greedy Randomized Adaptive Search Procedure (GRASP) to generate initial feasible solutions, a meta-heuristic à la Tabu Search (TS) to improve initial solutions, and other techniques based on the solution of reduced MILP problems. In the TS algorithm we proposed different types of moves (appliances shift, batteries charge or discharge…) to explore the neighbourhood. We have tested our methods on a data set of 180 realistic instances with different number of houses (20, 200 and 400), PV panels and batteries. The solutions provided by the heuristics are compared with those obtained by solving the MILP model by using a state-of-the-art solver. For instances without batteries all our heuristics yield high quality solutions – within 3% from the reference solution – in a short computing time for the largest instances. Heuristics that solve reduced MILPs achieved the same results even for instances with batteries