Solar photovoltaic (PV) modules are a popular source of clean energy, and efective monitoring and optimization of their performance requires the ability to explore and analyze the measurement data from their sensors. In this work, we introduce a tool for the visual analysis of such data stored in raw measurement files, ofering eficient interactive visualization directly on the files. The tool includes a soiling detection module, and novel UI modules that combine time series data visualization with advanced solar panel analytics presentation.
visual processing before users can start the visual analysis. This increases the deployment complexity and costs, The use of solar photovoltaic modules is becoming in- and limits the eficiency and the capabilities of ad hoc creasingly widespread as a way of generating clean en- interactive analysis directly over raw data files, that are ergy. Efective monitoring and optimization of the per- collected by sensor devices. formance of these modules require analyzing and visualizing the data collected from their sensors. This data, Contribution. The contributions of this demo paper are which may often be stored in raw data files (e.g. CSV as follows: (1) We present a tool for visual analysis of ifles), contains timestamped measurements of various sensor measurement data from solar photovoltaic modnumeric variables such as power output and solar irra- ules stored in time series files. (2) It ofers scalable visual diance. Such data are usually visualized on time series analytics capabilities without support of an underlying line charts and intuitive user operations are provided to database, through an in-memory multi-level aggregation assist their analysis (e.g. panning, zooming, filtering). tree that is built when the user starts a session, and ad
While there are many tools that provide generic time justs to user interaction for speeding up performance. (3)
series visualization functionality, they do not address the It ofers various advanced solar panel analytics including
specific needs of solar PV module data analysis. In par- a soiling detection visualization functionality. (4) Novel
ticular, the detection of significant events such as rain UI modules have been developed to combine time series
washes, and the determination of optimal times for man- data visualization with advanced solar panel analytics
ual cleaning to improve panel performance, are critical presentation.
objectives in the analysis of this data. Integrating these Related Work. Several techniques have been proposed
capabilities directly into the visualization would greatly for visually analyzing raw data files without the need
enhance users’ ability to analyze the data and optimize for loading or indexing in a DBMS [
0000-0002-9044-796X (V. Stamatopoulos); 0000-0003-2816-4368 (S. Maroulis); 0000-0003-3983-6449 (K. Kozanis); 0000-0002-5079-5003 (I. Psarros); 0000-0002-9273-9843 (G. Papastefanatos); 0000-0002-8252-9869 (G. Giannopoulos); 0000-0003-0784-8402 (M. Terrovitis)
© 2023 Copyright for this paper by its authors. Use permitted under Creative CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g CCoEmmUoRns LWiceonsrekAstthribouptionP4r.0oIncteerenadtiionnagl s(CC(CBYE4U.0)R.-WS.org) not require labeled data or generic analytical formulas, but instead relies on a minimal set of available measurements to estimate soiling with high accuracy. Although there are many tools for the visualization and analysis of time series data, to the best of our knowledge, this is the ifrst visual analytics tool specific for the visualization of PV module measurement data that also integrates soiling detection visualization functionality.
Exploration Scenario. The scenario we are examining involves a user who wants to visually analyze monitoring data from multiple PV modules in a solar park. The data consists of multivariate time-series that measure various metrics, including power output, irradiance, and temperature of PV modules, as well as weather parameters such as precipitation. This information is periodically collected and stored in data files 1, with each file containing measurements from a single module over a specific time period. The user wants to explore and visually analyze data for one or more PV modules using appropriate visualization methods, as well as interact with the visualization by panning, zooming, and filtering. Additionally, they intend to perform more advanced analysis to detect subpar performance of solar panels due to dust accumulation and schedule washings to improve their performance.
Architecture Overview. Fig. 1 illustrates the tool’s 1In our implementation, we consider row-based file formats and specifically CSV files. architecture. Measurements from each PV module are received in batches and stored in separate folders in the ifle system 1; . The user selects a module and defines a time period to analyze its measurements 2; . The data is parsed and stored in memory, in a tree-based structure, which is partially constructed based on the user request. During the construction of the tree, the results of the ifrst user request are also evaluated 3; . The user interacts with the visualization UI through a sequence of visual operations 4; , which are evaluated using the tree 5; . If necessary, missing measurements are fetched and the tree is updated 6; . If the user wants to perform soiling detection analysis, the sensor measurements are read and used to train a regression model and evaluate the soiling index 7; . The results are then sent back to the front-end for visualization. 8; Multi-level Aggregation Tree. The in-memory data structure is a multi-level aggregation tree which keeps at each level aggregated time series data from multiple files based on their timestamps. Each tree level represents a diferent temporal granularity (e.g. year, month, day, hour), coarser than the granularities of the levels below, and the finest level of granularity in the hierarchy is limited by the sampling frequency. Based on its path from the root, each node corresponds to a specific time period and aggregates the data points within it, storing aggregate values (e.g. mean, min, max, sum) over the time series variables. Since the measurements in the data ifle are ordered based on their timestamp, each node also stores the file ofset of its first data point and the total number of points to fetch the data eficiently, without having to scan the entire file.
user requests to visualize a PV module’s measurements by selecting the folder (where the files of the PV modules reside) and providing an initial time period of interest.
To speed up initial processing, we partially build the tree with data from the requested period and enrich it as the user interacts with the data. First, we read the first and last record from each file to determine their temporal range, and then only parse the files with measurements within the time period requested to construct a partial tree. The first request also determines the maximum level of temporal granularity of the initial tree, as well as the level of details for diferent objects in the data. For instance, if the user requests data for every minute of the last 24 hours, those measurements will be aggregated up to the minute level, while the remaining data will be aggregated at coarser levels of detail.
After initializing the tree, user operations are evaluated by traversing the tree up to the granularity level specified in the query, considering only nodes overlapping with the query time interval. If a node is found that corresponds to the user’s requested frequency, its timestamp and aggregate values are returned. If parts of the time series are missing or not pre-aggregated for the requested frequency, the required measurements are read from the corresponding files, aggregated, and the tree is enriched with finer granularity sub-trees. able on GitHub 2, features a front-end developed with TypeScript and React, designed as a single-page application. The back-end, written in Java 11, interacts with the front-end through a REST API.
allows users to select a specific PV module by selecting it from a list of available modules and visually explore its data on the chart A . They can analyze the sensor data in a sampling frequency of their choice B and select to visualize diferent sensor variables from the Measures Panel on the left C . Also, they can temporally explore the data by panning left or right, or zooming in and out on the chart. To speed up the visualization process on large data, we utilize the in memory tree described in Section 2, whenever possible, instead of reading the raw data files.
able to perform several practical operations on the data depicted on the chart. These operations are enumerated Soiling Detection Module. The Soiling Detection mod- from left to right. a) Manual selection of the visualized ule uses as input the PV module measurements within date range; b) compare sensor data with another PV the user-specified time period and needs to fetch the cor- module in the farm. For example, the user can visualize responding data points from the disk. It may also take the power of two diferent PV modules, on the same chart, as input dates during which the solar park was manually to detect potential under-performance issues; c) select cleaned as potential cleanings, along with raining periods. between diferent chart types (line, candle); Finally, d) The module first determines efective cleanings, i.e. the the user can change the chart view from stacked, where ones with positive efect on energy production. To that the sensor variables are visualized on diferent charts; end, it checks each potential cleaning for the existence of to overlay, where sensor variables are visualized in the a changepoint in the behavior of power output as a func- same chart. tion of irradiance and PV module temperature. This is Soiling Detection. The PV Dashboard ofers a toolkit implemented as follows: a regression model is trained be- that provides the user with a set of tools, to help them fore the potential cleaning and tested after the potential detect interesting patterns in the sensor data. One such cleaning; prediction errors are then compared to deter- tool, is the Soiling Detection tool E , which includes a mine a change of behavior. We apply Ridge Regression toggle to visualize the manual washes on the chart or the with polynomial features. Our choice satisfies a two-fold potential rains that acted as washing events, and were objective: i) good accuracy and ii) fast fitting time. The detected by the Soiling Detection Module. The user can latter is vital in our method which fits one model for also specify in the chart intervals in which they know each potential cleaning. The method proceeds by train- that a certain washing event took place. The various ing a regression model on periods following the efective washing event intervals are highlighted with diferent cleanings. This model aims to capture the optimal ex- colors based on their type as shown in Fig. 2 (rains are pected performance, i.e., power output when the solar colored blue, manual washing events with green, and panels are clean. Finally, the Soiling Index is returned, user-highlighted washes with yellow). which is a performance index calculated by dividing the To help users analyze washing events detected by the actual power output by the expected power output for algorithm, a slider can be used to filter events based on the whole time period.
their score. When the user enables event visualization, amount of solar irradiance in watts per square meter rethe visualization changes and displays the power and ceived on the module’s surface, the modules back-surface precipitation sensor measurements. The Soiling Detec- temperature and the accumulated daily total precipitation Module can then be enabled with input parameters tion. including the washing events and the time range the Users will be able to interact with the prototype and panels are considered clean after each wash. Finally, the perform several operations such as: a) Interact with the Soiling Index is displayed on top of the chart, allowing chart to pan, zoom in/out to find areas of interest and users to view estimated power loss per measurement or filter the visualized datapoints. b) Select diferent meaaggregated power loss in a specific time range. sures (sensor data) from the Measures Panel. c) Visualize Washing Events and apply the Soiling Detection module’s Filtering. By utilizing this tool, users can specify range algorithm on these events. iflters for the sensor variables, and easily filter out any Users will also perform specific analytical tasks to betdata points that fall outside those ranges F . When a ter understand the tool’s capabilities. For instance, if iflter is applied to a certain variable, the corresponding a solar park operator wants to determine the best time measurements are filtered from the visualization of all to perform a manual wash of a solar panel to minimize variables. For example, a user may want to visualize the power loss due to soiling, they can perform Soiling Depower of a PV module, where the irradiance in a specified tection on one of the solar panel datasets. Initially, they period was above a certain threshold. This way, they can can visualize the relevant period of interest on the chart easily distinguish outliers, i.e. data points with low power by performing pan and zoom operations, and select the despite the high irradiance. This, in combination with the types of washing events to view in the Soiling Detection Soiling Index visualization, can help them spot anomalies tool. Then, they can enable Soiling Detection to visualize in the PV module’s activities. the Soiling Index above the chart. Using this information, they can analyze how each washing event afected the 4. Demonstration Outline power loss of the panel. For example, if the Soiling Index increases after a washing event, the user can infer that In this section, we outline our demonstration scenario. the washing was efective.
The tool is available at: Subsequently, the user can filter the detected washing http://visualize.more2020.eu/visualize/solar/eugene. events to keep only the most significant rains, and detect The attendees will be able to interact with the tool and their parameters, such as duration and periodicity, that analyze Solar Panel sensor data. The main focus of the contributed to the cleaning of the panel. This information demonstration is the Soiling Detection module and the can be used to schedule future manual washes of the visualization of its results as described in Section 3. panel. For example, if the latest Soiling Index is low and
The datasets used for the visualization are provided a similar washing event is not expected in the near future,
in [