1. Introduction
Renewable energy sources (RES) such as the photovoltaic (PV) system have played an important part in reducing environmental pollution in recent years due to their ability to reduce greenhouse effects [
1]. As a kind of mature and widely used power generation method, PV power generation perfectly conforms to the strategy of sustainable development and the concept of safe power generation. With the development of distributed generations (DGs), PV can be operated at a smaller scale called distributed energy resources (DER). This form of PV aims to be closer to the load that needs to consume power, which uses the idea of decentralized investment to reduce the loss in the transmission [
2]. However, PV has the characteristics of intermittence and instability. Solar irradiation, cloud cover, photovoltaic panel orientation or dust diffusion and other factors may interfere with the normalized PV to a great extent [
3]. Besides, high penetration of PV may lead to the issues of voltage rise, reverse power flow, and increased energy loss [
4].
Nevertheless, with the rapid development of electric vehicles (EVs) in recent years, it is becoming increasingly possible to alleviate this negative phenomenon. Some studies show that EVs can not only carry out routine charging operation, but also profitably provide power to the power grid after the EVs with new technology called vehicle-to-grid (V2G) are connected to the power supply [
5]. Therefore, a new idea for implementing energy storages to the grid which has a strong impact on the traditional distribution network, and EVs are gradually being accepted by people [
6]. V2G greatly improves the flexibility and availability of EVs. In general, PV is greatly affected by time changes. The energy storage of EVs can solve the problems of power overflow and voltage rise caused by PV at noon, and realize the operation of power transmission to the power grid at the peak of power consumption at night [
7]. In addition, in response to the worldwide appeal to reduce carbon emissions, the use of EVs is expected to become more widespread. This trend caters to the need to solve the problems caused by new energy generation in the distribution network.
Some research has shown that when distributed PV and EVs are connected to the distribution network, reactive power optimization of distribution system will be a complex discrete, nonconvex and nonlinear problem [
8]. In the daily operation of the distribution network, line loss of transmission and voltage deviation are usually the issues that need attention. It is necessary to optimize the variables with line loss minimization and voltage deviation minimization as two different objectives. The classical methods commonly used by researchers to solve such multi-objective problems (MOPs) are the interior point method and the Newton method. The idea of these methods is to transform the MOPs into some single objective optimization problems with assignable weight. The problem highlighted by these methods is that the setting of weight is highly subjective and may not reach a balance point. Using the multi-objective evolution algorithm (MOEA) to solve MOPs is also a good choice for current optimization problems, but these methods usually spend a lot of time on optimization in the face of complex problems.
The characteristics of PV and EVs show the ability to provide reactive power support for a distribution network. As a premise, the future development trend of different types of EVs was well estimated in reference [
9]. With the combination of a residential roof PV system and EVs, the authors in [
10] demonstrated the potential and technical benefits of such system in terms of the reduction in air pollutant emissions. In reference [
11], the robust dynamic evolutionary optimization of the reactive power system in interconnected systems under fluctuating and uncertain wind power conditions was proposed. It has been suggested that leveraging the reactive power range embedded in wind farms can improve safety and optimality during the power system reactive power optimization process [
12]. The reactive power optimization including interval uncertainty model was applied for developing a voltage control strategy to ensure that the state variables of a power grid reside within their safe operating limits [
13]. The authors of [
14] considered the hydrogen and PV as distributed generation and proposed a list of reactive power regulation strategies. However, the work presented above does not mention the participation of PV and EV in reactive power optimization.
Some researchers have synthesized the active power generation and consumption as constraints and considered the cost of reactive power injection. The particle swarm optimization (PSO) was used to optimize the reactive power [
2]. Similarly, aiming at the cost, the authors of [
15] presented an in-depth study on the PV-Biomass hybrid independent power generation system in remote areas. Non-dominated sorting genetic algorithms III (NSGA-III) were proposed to address the reactive power optimization model which was established with the objective of minimizing system active power losses, controllable loads reduction, and PV active power reduction [
16]. The authors of [
17] discussed the reactive power based on capacitors allocation by using mathematical remora optimization algorithm. A data-driven model was employed to address the uncertain output of distributed generators for reactive power optimization in reference [
18]. In reference [
19], reactive power, the number of shunt capacitors and transformer taps were taken as optimization objectives and solved by the improved firefly algorithm. The authors of [
20] pointed out that the reactive power control and actual power reduction of photovoltaic inverter could be effectively solved by a global sequential quadratic programming (SQP) approach method. Despite all this, these algorithms manifested the high consumption in time and instability in optimization.
By contrast, as a popular technology in recent years, deep learning (DL) may replace the traditional optimization process in the aspects of fitting data and improving the efficiency. The biggest feature of this method is that it can use a large amount of data for supervised learning and predict the optimization results in other cases via the previous optimization data set [
21,
22,
23]. In this paper, five different multi-objective optimization algorithms are listed, and the same problem is optimized respectively. Through the training of DL, the optimization results of the five algorithms can fit the Pareto front (PF) like the previous optimization process. The fitted PF can be further optimized by correcting and supplementing some negative points.
Thus, the innovations of the proposed technique for reactive power optimization in this paper can be summarized as follows:
In general, the reactive power optimization only considers the regulation of traditional equipment without the participation of PV systems or EVs, so that the reactive power regulation capacities of these new regulation sources are wasted. In this work, PV and EVs are simultaneously employed to participate in reactive power optimization in a distribution network, which can greatly decline the pressure of traditional reactive power regulation and improve the regulation flexibility and performance.
To address the multi-objective reactive power optimization, the meta-heuristic based Pareto optimization algorithms easily result in a long computation time to acquire the high-quality Pareto optimal solutions. Besides, they easily lead to different Pareto front in different runs due to their random operators. In contrast, the proposed deep learning-based Pareto optimization algorithm can acquire the high-quality Pareto optimal solutions within a short computation time since it cannot experience multiple iterative operators. Moreover, it is a deterministic algorithm to guarantee a high optimization stability.
The rest of this paper is structured as follows. In
Section 2, the model of PV and EVs connected to the distribution network will be explained.
Section 3 will illustrate different algorithms simply which are applied on the model, and introduce the flow chart of the reactive power optimization.
Section 4 presents the experimental results and analysis for different examples. Finally, the work will be concluded in
Section 5.
5. Conclusions
In this work, a novel DL based Pareto optimization method is proposed for multi-objective reactive power optimization in a distribution network with PV and EVs, which contains the following contributions:
(1) By taking the participation of PV and EVs into account, the reactive regulation burden of the distribution network can be effectively reduced. As a result, the operation economy and the voltage quality of the distribution network can be further improved. Simulation results demonstrate that the line loss can be reduced by 25.2% by the proposed method compared to that without the participation of PV and EVs on the IEEE 14-bus system, and 7.7% on the IEEE 33-bus system. In addition, the voltage deviation can be reduced by 0.16% and 0.38% on the IEEE 14-bus and IEEE 33-bus systems, respectively.
(2) The proposed technique for multi-objective reactive power optimization is verified on IEEE 14-bus and IEEE 33-bus systems, which is compared with five different intelligent algorithms. Simulation results show that the proposed DL method can rapidly acquire a high-quality PF, while another five algorithms can complete the optimization task well, and the Pareto optimal solutions have little difference in the degree of optimization. Particularly, the computation time of DL is only 1.26% of that by NSGA-II on the IEEE 33-bus system, while the average line loss and voltage deviation of all the Pareto optimal solutions obtained by DL are the smallest among all the algorithms.
Although the proposed DL based multi-objective reactive power optimization can perform well on the test systems, it still easily faces two main challenges. Firstly, it should spend a long computation time on the data acquisition and network training. Secondly, it will easily lead to multiple infeasible solutions due to the poor generalization of DL with insufficient data. To handle these two problems, our future work will focus on the historical data utilization and the generalization for DL.