As a robot navigates in an indoor environment, it can collect a dataset of local observations using the on-board laser range-finder, which can then be used to train machine learning models for place classification. The goal of our project is to enable storing images of laser-range observations into a relational database, and retrieving similar images to a query image efficiently. To this end, we have developed a pipeline to extract image features and store them into a relational database, following the Bag-of-Visual-Words model. Then, we implemented FIDS ("Flexible Image Database System") to speed up image retrieval, as well as Locality Sensitive Hashing to store data in a way that allows for faster retrieval. Our experiment results show that both methods are significantly more efficient than brute force, and do not cause a noticeable decrease in retrieval accuracy.

Solar irradiance forecast is the pivotal factor for power generation scheduling at photovoltaic (PV) power plants. To this end, we propose a framework to model solar irradiance that leverages both short-term dependency (several days) and dependency on time in the annual cycle. We focus on two classes of methods, autoregressive (AR), and autoregressive with exogenous variables (ARX), and investigate LASSO regression, and non-linear methods, including support vector regression (SVR) and long short-term memory (LSTM) as the underlying mechanism. Our experiments on data collected at a PV power plant for over five years show that considering external information such as weather on top of the historical solar irradiance data for longer period of time lowers the error of solar irradiance forecast.