| With the deepening of the power system reform,the power market supervision mechanism is not perfect,and the collusion credit risk of the power market subject is frequent,which seriously damages the fair and healthy market order.At present,the scope of access to the power market is expanding,the number of market subjects is increasing,the scale of power transactions is getting bigger,and the spot transactions are becoming more frequent,so it is definitely not feasible to rely on expert decision-making for early warning identification of market collusion.Therefore,in the face of massive market data,a data-driven machine learning approach is used to dynamically simulate the bidding decisions of generating companies and to provide intelligent real-time early warning of collusive behaviors among them.The main work of this paper is as follows.(1)A set of bidding decision models for generators based on Twin Delayed Deep Deterministic Policy Gradient(TD3)agent is proposed to analyze the bidding behaviors of generators under greed/non-greed.Firstly,a bidding model and a market clearing model are constructed based on the trading means of generators,and the bidding process is described as an infinite repetitive game process under limited information.Secondly,the current state of deep reinforcement learning and representative algorithms are introduced,pointing out that the intelligences can continuously interact with the market to optimize their bidding strategies and converge to Nash equilibrium without greed(discount factor of 0).Then,the superiority of the TD3 over other representative algorithms is verified in the arithmetic example of the 3-bus system,which can converge to the optimal policy quickly and stably in the process of policy optimization for generators in the electricity market.Finally,the bidding behavior of generators under greedy situations is analyzed to reveal the relationship between the revenue discount factor of power generators and tacit collusion.(2)Combining the collusion early warning index system of generation firms and the unsupervised Variational Autoencoding Gaussian Mixture Model(VAEGMM),the intelligent early warning of collusion is realized.Firstly,a complete collusion early warning indicator system is constructed,and the labeled indicators and the measurement methods are proposed,pointing out that the indicator set has the data characteristics of high-dimensional and unbalanced positive and negative samples.Secondly,a new unsupervised learning algorithm VAEGMM is proposed for the data characteristics combined with the idea of anomaly detection.Then,the network structure of VAEGMM is described in detail,and the joint loss function is derived.It shows that the network can better learn the low dimensional expression of the original data and it is helpful for more accurate density estimation.Finally,it is verified that the VAEGMM can efficiently warn the collusion risk of power generation companies in both thresholded and unthresholded cases,and its superiority over other unsupervised learning models and collusion identification methods is illustrated.(3)Based on the mentioned collusion intelligence warning methods,a collusion intelligence warning system for the power market is developed.The practical application scenarios of collusion early warning are discussed,and it is pointed out that the method can improve the supervision module of the future power big data trading platform to monitor the market credit risk in real time.Then the hardware and software environment and specific functions of the system are introduced in detail,and the future direction of software development is proposed. |
PDF Full Text Request