Errors in the reading and measurement of the energy generation of photovoltaic plants are a major problem when it comes to boosting the performance of the installations.
At Smarkia we develop disruptive solutions both at a technical and scientific level to achieve an energy transition towards a more sustainable system. That is why we have developed a solution for this problem of measurement of photovoltaic plants. It is a research based on the use of Artificial Intelligence (AI), within the branch of study of the Deep Learningand has been published in Expert Systems With Applicationsan international reference journal in its scientific category.
In this article we provide you with a summary of the scientific paper to show you how Deep-Learning improves data quality in photovoltaic plants.
What is Deep Learning?
First of all, we would like to clarify this concept. The term Deep Learning refers to a specific field of study and constitutes one of the most researched branches of Artificial Intelligence at present. This branch of study encompasses all artificial neural networks that have depth. These neural networks are composed of several layers and enable the most complex AI operations to be carried out.
The problem of PV generation data loss.
Solar photovoltaic production is growing exponentially in Spain. According to data from REE (Red Eléctrica de España), in December 2022 the national photovoltaic generation park was close to 19,000 megawatts of power. In other words, it has grown by 3.8% in the last year alone.
This increase in photovoltaic production brings numerous benefits to society as a whole, but it also presents several technological challenges. One of these is the lack of ability to accurately measure energy production and its relationship to other factors.
This problem has a greater impact than might be expected, since it affects both the operation and maintenance of the installations, as well as the forecasts that are necessary to correctly integrate solar energy into the electricity system.
Operational and maintenance problems caused by the absence of data.
Handling reliable and accurate data is key to perform a correct analysis of production and advanced predictive maintenance, which results in greater profitability and efficiency of the installation.
These data are necessary to understand and predict the plant's electrical generation over time, as well as to improve its performance and useful life.
Production forecasting problems associated with data loss.
Photovoltaic production depends on meteorological factors such as solar radiation, temperature, relative humidity or precipitation, among others, which causes fluctuations in production. These fluctuations can be a problem in grid-connected installations, as a certain predictability is necessary to integrate energy into the power system operator.
Therefore, the ability to make accurate forecasts is essential for the integration of photovoltaic energy. The problem lies in the lack of production data, weather data, etc., in many of the series used to make these predictions, which negatively affects the quality of the forecasting and, therefore, the operability of these plants.
Why does data loss occur?
Data loss in photovoltaic production is a common occurrence and is due to a multitude of factors such as:
- Connection/communication problems
- Shutdowns for maintenance or repair of inverters
- Occurrence of outliers caused by noise
- Malfunctioning due to atmospheric phenomena
- Sensor problems
- Infrastructure problems
- Etc.
Some authors indicate that a time series with a loss of up to 5% of the data already requires treatment before it can be used. In many cases, data loss can reach up to 40% of the solar production information, which prevents these data series from being used for analysis.
In these cases it is necessary to perform an estimation of missing data to enable a comprehensive analysis on them. This problem has been studied before, although it has only found a solution through the application of AI that exceeds previous achievements. Smarkia has found a solution through the application of AI that surpasses previous achievements.
How AI can solve the problem of missing measurements.
The application of Artificial Intelligence and Deep learning has already been used to solve classification and prediction modeling problems in other fields, such as the diagnosis of some types of cancer or the prediction of traffic problems.
However, the innovation of Smarkia is the application of an artificial neural network that can be trained on a single set of energy production data. Unlike other techniques, this model learns the variability patterns present in the data and predicts accurately even when the same data set is used for training. This offers an improved solution for missing data imputation.
In addition, the model can work with data series in which up to 70% of the data has been lost, although the best results are achieved when this value is around 50%, which exceeds the figures achieved by other models to date.
Benefits of AI application in the measurement of energy production and storage.
The model proposed in the study allows the completion of partial data sets, thus facilitating the operation and forecasting of electricity production in photovoltaic plants. Additionally, this new approach manages to do so by improving what other models had achieved to date, reducing the need for previous data, improving the processing speed and the quality of the data obtained.
Less need for previous data.
This model is able to learn from the data in an unsupervised way without the need to be trained with different example data. In fact, the model can be trained with the same dataset to be reconstructed, making both training and prediction part of the imputation process for the same series.
The advantage of this solution is that, although it is a Deep-Learning model, it works as soon as there is a single time series for training, even if it is incomplete. In other words, it is trained very quickly without the need for a previous data set.
Improved processing speed.
Although the proposed method is based on Deep-Learning, meaning that it has several layers with adjustable parameters to fit the data, the model can generalize the data prediction of missing values with as little training as that performed with a single partial sample.
This is true even if the sample is incomplete and must be reconstructed, which makes the training and prediction stages part of the same imputation process, thus reducing processing times.
In addition, no other external correlated signals are required to train the model, making it ideal for restricted contexts lacking prior data or in environments where correlated signals such as solar radiation, temperature or wind speed are not known.
Improved data quality.
Our solution is capable of acquiring data variability factors from a single example sample (these factors represent how the data varies over time), for example, due to the difference in solar radiation throughout the year, the variation in sunrise and sunset times, or the angle of radiation, among others.
In this way, the prediction of the missing data respects the distribution of the existing data in the series and, therefore, predictions that are more in line with reality are achieved.
SmarkiaInnovative and state-of-the-art technology.
At Smarkia we are looking for new ways to optimize the energy transition, and we are not satisfied with established solutions. The publication of this paper is another example of our cutting-edge technology and innovative capacity, which allows us to operate with new and better solutions to the challenges of the present and the future.
If you want to access the full paper you can do it here.
From Smarkia we put our proven excellence at the service of companies that want to lead the transition to a sustainable future through non-conformist, disruptive and innovative solutions that offer tangible results. If you want to join the energy(r)evolution you can contact us through our website.
