A fully decentralized machine learning algorithm for optimal power flow with cooperative information exchange

Abstract

Traditional power grids, being highly centralized in terms of generation, economy, and operation, continually employed probabilistic methods to account for load uncertainties. In modern smart grids (SG), the rapid proliferation of non-dispatchable generation (physical decentralization) and liberal markets (market decentralization) leads to the dismantling of the centralized paradigm, with operations being performed by several decentralized agents. Handling uncertainty in this new paradigm is aggravated due to 1) a vastly increased number of uncertainty sources, and 2) decentralized agents only have access to local data and limited information on other parts of the grid. A major problem identified in modern and future SGs is the need for fully decentralized optimal operation techniques that are computationally efficient, highly accurate, and do not jeopardize the data privacy and security of individual agents. Machine learning (ML) techniques, being successors to traditional probabilistic methods are identified as a solution to this problem. In this paper, a conceptual model is constructed for the transition from a fully centralized operation of an SG to a decentralized one, proposing the transition scheme between the two paradigms. A novel ML algorithm for fully decentralized operation is proposed, formulated, implemented, and tested. The proposed algorithm relies solely on local historical data for local agents to accurately predict their optimal control actions without knowledge of the physical system model or access to historical data of other agents. The capability of cloud-based cooperative information exchange was augmented through a new concept of s-index activation codes, being encoded vectors shared between agents to improve their operation without sharing raw information. The algorithm is tested on a modified IEEE 24-bus test system and synthetically generates historical data based on typical load profiles. A week-ahead high-resolution (15-minute) fully decentralized operation case is tested. The algorithm is shown to guarantee less than 0.1% error compared to a centralized solution and to outperform a neural network (NN). The algorithm is exceptionally accurate while being highly computationally efficient and has great potential as a versatile model for the fully decentralized operation of SGs.