This Small Business Innovation Research Phase I project seeks to develop new methods and algorithms for the analysis and control of degradations in Photovoltaic (PV) systems. There currently exists widespread, and often undetected, underperformance in PV arrays. Although there are numerous site evaluation and analysis tools used in the design phase of solar cell installations, there is currently a lack of on-line monitoring and diagnostic tools to detect and estimate underperformance and associated control algorithms that can improve performance. Underperforming PV cells can result from a host of inter-related abnormalities including shading, dust build up, hot spots and component failures. To improve performance intelligent control can be used to ameliorate power output degradation via actuation at the module- or module-string-level operating points. The purpose of this proposal is to develop, and test in simulation, a hybrid MAP-MPPT estimation and control algorithm, capable of intelligently modulating the operating point to maximize
The broader impact/commercial potential of this project include the enhancement of performance of large scale PV arrays and the ability to better estimate the energy payback time for commercial and residential PV systems enabling further deployment of PV worldwide. Leading revenue-critical, PV power plant owner-operators often utilize dedicated operation centers to monitor walls of computer screens, sometimes with alarm mechanisms triggered by inverters. It is often the case that ad-hoc methods are utilized and decisions are based primarily on operator experience and intuition. Although some sites utilize commercial web-based monitoring services, with industrial grade sensors, this process is far from rigorous, and has resulted in only spotty success. Furthermore, the lack of intelligently controlling the operating point has not yet been fully developed. The proposed solution, used in conjunction with the ability to monitor the state of the PV plant, will correct faults via intelligent control methods enabling optimal power production under degradations for large scale PV arrays. Potential customers are revenue critical solar power sites including: Power Purchase Agreement (PPA) contractors, and lease-holder hosts.
’. In this project, we have logged thermal, electric, optical irradiance and wind-speed sensor data from our roof-top PV array setup, using a single panel at a time (one of three types), each comprising a single multi-group string of photovoltaic cells, with a bypass diode across each group, and a variety of both naturally-occurring and artificially-induced non-uniform environmental irradiance patterns. In addition to the empirical data, we have also generated simulated photovoltaic data, using our integrated optical-electrical-thermal PV plant model, applied to the same array configuration, with the baseline parameters of the simulated array obtained via system identification from actual array data. For the simulations, our existing PV plant model was augmented to incorporate PWM controlled DC-DC converters (at various granularity levels), which were not present in our actual array. The simulated sensor data, in conjunction with a machine-learning diagnostics algorithm suite we have developed and our augmented model for the controlled array, were used to generate PWM control signals for the converter actuators which effect Maximal Power-Point Tracking (MPPT). Our MPPT approach can be used to ameliorate the power reduction, accelerated hot-spot caused degradation, and monetary costs of varying array mismatches caused by unknown non-uniformity patterns in the irradiance impinging upon the PV array. Our machine-learning diagnostics algorithm suite consists of three components: (I) Trained Neural Networks for cell I-V characteristics Estimation and MPPT: (II) A System-ID algorithm which extracts from offline PV cell data a simplified (Spline) parameterized mathematical model of I-V electrical characteristics at the cell- and coarser granularity levels; (III) A Bayesian/MAP (Maximum Aposteriori Probability) algorithm for detecting and estimating hot-spot regions in the photovoltaic array; attributing them to proximate and root causes, such as electrical mismatches and irradiance non-uniformities; and prognosticating the future evolution and performance consequences of these hot-spot degradation events. The actuation method we used in our Phase I augmented PV plant simulation model is a DC-DC buck/boost converter; we explored different intelligent control granularity levels at which to place converters; for an example topology of a controlled array. Each DC-DC converter has two input leads, two output leads, and an input control voltage signal for PWM control of its gain. The internal operation of an actual such converter involves an intermediate AC stage, inductances and a (e.g. MOSFET) switch. In the Phase I effort, however, we circumvented such details by using simple effective (‘behavioral’) models for the overall Kirchhoff laws of each converter. In the simplest such model, the output-to-input voltage ratio is reciprocal to the corresponding current ratio, and proportional to the PWM control voltage signal. The goal of the machine-learning diagnostics algrorithm suite described above, is then to (I) use partial sensor data from the entire PV array (and optionally from cloud sources, e.g. weather data) to detect and estimate the degradations (primarily incipient runaway hot-spots) and their environmental causes at any given moment, and prognosticate their future evolution (possibly leading to irreversible component damage); and (II) to construct a virtual search of the OP (operating point) landscape of the array, and determine at any moment an optimal PWM control command to be inputted into the local DC-DC converter actuators, with the goal of maximizing the overall power provided by the degraded array. A secondary goal of the machine-learning algorithm suite was (III) to optimize — from the standpoint of monetary maintenance cost and hits in array energy output — the ensuing maintenance scheduling.
