Issue link: https://iconnect007.uberflip.com/i/1104607
24 PCB007 MAGAZINE I APRIL 2019 talk to its neighbors, and there is a relay con- nected to every neighbor. The connected relay means you can change the connections based on whether a particular panel is not working well, for example. If one panel is not efficient, you can bypass that panel. That's how you can change the topologies we can make connec- tions between a different number of panels. The questions that follow include how do you change those panels? How do you know when the topology is going to change? And when are you required to change the topology? One way is to study the panel's output over a period of time. You might see, for instance, that a particular panel didn't work up to its optimum. But that doesn't serve our purpose because already we've wasted a huge amount of time reading the panel values, so we try to automate this process. The first thing we try to do is we have a cloud movement prediction algorithm. We use an ap- proach called nowcasting, which predicts at what speed and in which direction the clouds are going to move. Based on that shading, we can change the topology. For example, a par- ticular section of panels could become shaded, which means the power output will go down. We can bypass those panels. Johnson: So, based on weather conditions that you have monitored and studied in the past, you have a predictive awareness of what's go- ing to happen tomorrow and can be roughly prepared for it? Katoch: Yes. Now, different faults need differ- ent kinds of topology. You have to identify a panel as faulty so that you can bypass it, and you cannot do it if you don't know whether the panel is faulty. To determine that, you use many automated ways, such as AI clustering techniques. That is an unsupervised machine learning technique with the two basic faults— ground and arc faults—unlike the normal pan- els. We have to identify three clusters of ground faults, normal working panels, and arc faults. In the case of shading and soiling, you see a similar I-V curve, only the voltage at maximum power (VMP) and current at maximum power (IMP)decrease. Clusters will form around the same region, and that's where the system cannot differen- tiate. To solve this problem, we switch from unsupervised to supervised techniques. That's why we use a fully connected neural network. At this point, we have a training data set. And in the training data set, we have the relation- ship defined. For example, this particular kind of data gets generated when there's a ground fault, or this other kind of data is generated when there's an arc fault, shading, or any oth- er faults. Once the model has that set of conditions de- fined, based on the input features, the neural network output gives the type of fault. Then, we try to minimize the difference between the maximum power output from that vicinity to the power output that is generated. When you optimize this loss, the weights between the neu- rons are learned. Once the weights are learned, when new input comes in, these weights help identify which fault is causing the change. That's the basic idea. Emma has characterized three categories of outputs, which are scalable to much more faults. We also submitted papers and patents on this research. Johnson: A lot of these fundamental concepts must be transferrable. For example, I have a per- sonal interest in sailboats, which can be dynam- ic platforms and could easily stress solar panel Sameeksha Katoch with winning poster.