شبکه عصبی مصنوعی مبتنی بر ردیابی نقطه حداکثر توان برای شرایط متغیر سایه

ABSTRACT ANN based MPPT method for rapidly variable shading conditions

As it is well known, the use of photovoltaic systems (PV) are becoming more and more important due to their environment friendly and economically sustainable energy source, also in electric vehicles. On the other hand, the photovoltaic technology still faces efficiency limits, thus control techniques named Maximum Power Point Tracking (MPPT) have been proposed to optimally exploit the available power. These algorithms are tracking controls employed to extract the maximum power from PV modules depending on the array temperature, solar irradiation, shading conditions and PV cell ageing.

The most widely used MPPT methods can be grouped in two different categories: hill climbing methods, such as Perturb and Observer (P&O) and Incremental Conductance (INC), and constant voltage methods [1–4]. Starting from the standard implementations, other technical solutions [5–18] have been proposed in order to improve the accuracy and dynamic behavior of the tracking controls. On the other hand, most of them neglect that MPPT is a multimodal optimization problem [19] since there are local optima in the P–V characteristic curve when not uniform irradiance occurs over the photovoltaic system.

Considerable research efforts have been directed toward the development of more sophisticated MPPT algorithms able to identify the Global Maximum Power Point (GMPP) in order to extract the whole available power from the PV system under partially shaded conditions [20–33]. The computational burden, range of effectiveness and convergent speed of these algorithms is quite different and depends on the adopted theoretical methodology. Some methods determine the GMPP by exploiting deterministic searching algorithms such as constant power operation [21], dividing rectangles (DIRECT) method [22], restricted voltage window search algorithm [25], Cuckoo search [30], while other MPPT algorithms are based on metaheuristic approaches such as the particle swarm optimization [20] and Artificial Neural Networks (ANNs) [26]. More in general, in the field of multimodal optimization the Evolutionary Algorithms [34], e.g. genetic algorithms [35], applying niching strategies [36] are designed to properly address multimodal functions. Furthermore, some niching algorithms are coupled with Deterministic Methods (DMs) [37] to enhance the final solution accuracy [38]. The basic idea is to firstly search the global optimum using a niching algorithm and, then the DM starts from the provided solution to draw up to the actual optimum. Therefore, a further distinction among algorithms that can be adopted to MPPT concerns the number of stages that the MPPT algorithms use. In fact, some techniques track the GMPP by using a unique procedure, while other methods identify the GMPP by adopting two stages. The latter ones, firstly adopt an algorithm to identify the ‘‘hill’’ where the GMPP is potentially located, and then a further algorithm is employed to reach the GMPP. In this perspective, the use of metaheuristic approaches coupled with a DM seems a good solution provided that information about the solar irradiance distribution on the panels and/or their temperature as well as the knowledge of the system model are given. Actually, these techniques could be applied without the use of a model, but the approach becomes very harmful because each objective function evaluation calls for changing the PV voltage in order to measure the current.

Hence, the techniques mentioned so far require the measurement of solar radiation over the panel and their temperature, and/or the scansion of a large portion of the PV characteristic to suitably determine the GMPP. Consequently, the main limitation of the first kind of techniques is the necessity of using additional sensors and properly system models, whereas the main limitation of the second ones is the lost of energy owing to the time spent to sample the PV characteristic. In this perspective, in [39] a techniques based on the DIRECT search algorithm (first stage) and a suitable P&O algorithm (second stage) has been proposed. More in general, the aforementioned techniques have been designed for PV system placed in fixed installation, where the shading phenomena does not suddenly and frequently change as in case of installations on the roof of electric vehicles. In this case, a very accurate and fast GMPP tracking is necessary in order to maximize the extracted energy, taking into account that the PV system operates under the high probability that the solar irradiance on the panel is not uniformly distributed, especially due to the presence of other vehicles, buildings and any other obstacles that blocks or refracts the solar rays impacting the PV modules, and this distribution continuously changes meanwhile the vehicle moves in the traffic.

In order to overcome the aforementioned limitations, the paper aims to study the effectiveness of an ANN based MPPT approach whose goal is to quickly and accurately estimate the GMPP when no information about the solar irradiance distribution over the modules and their temperature is given, and when the PV system is subjected to continuously and rapidly changing shadowing patterns. Few measures are used to esteem the GMPP by means of the ANN, and they are set a priori. Consequently, the estimation time is small and fixed. In this work, the solution provided by the ANN is given, as starting point, to a zero-order P&O method, named as Pattern Search (PS) [37], in order to improve the accuracy of the esteemed maximum power point (EMPP). Therefore, differently from [39] the ANN is used to directly estimate the GMPP at first stage, while the P&O method is exploited only to refine the result. Finally, to evaluate the performance of various ANN’s structures some quality indices are proposed, and the robustness of the proposed approach to parameter variations of PV system has been also assessed.

Actually, ANNs have been already applied to MPPT problems, but it is the first time that ANNs are used to identify the GMPP as proposed in this paper, far as authors know. The other ANN based MPPT algorithms presented in past literature are usually exploited to compensate the parameter variations occurring in the PV system in the identification of local MPPs [14–18]. On the other hand, the ANN based MPPT method presented in [32] estimates the GMPP by means of a single voltage and current measurement but the study is limited to a very restricted number of shading combinations, thus it is few practicable for a generic PV system especially when it is continuously subjected to very fast changing and not uniform shading conditions.

The paper is structured as follows: a brief description of the considered PV model is presented in Section 2, while Section 3 deals with the explanation of the entire proposed MPPT method; a detailed analysis of the ANN based MPPT method is described in Section 4, and a case study is investigated in Section 5; finally, some conclusions and future work are pointed out in Section 6.