Researchers in Greece have developed three completely different transfer-learning methods for solar energy forecasting. The methods are based mostly on a stacked lengthy short-term reminiscence mannequin, which is a type of recurrent neural community that may study order dependence in sequence prediction issues.
Scientists on the Nationwide Technical College of Athens have used a machine-learning technique referred to as transfer- studying (TL) to develop a brand new solar energy forecast modeling meant at serving to builders with restricted information dimension.
The TL technique makes use of a educated mannequin on one process to repurpose in a second, associated process. The researchers used three TL methods together with the stacked lengthy short-term reminiscence (LSTM) mannequin, which is a type of recurrent neural community able to studying order dependence in sequence prediction issues. The LSTM method takes the related elements of a pre-trained machine studying mannequin and applies it to a brand new however related downside.
“TL is exploited each for weight initialization of the LSTM mannequin and for function extraction, utilizing completely different freezing approaches,” they defined. “LSTM will depend on weight updating between the neurons of the deep studying mannequin, permitting the creation of pre-trained fashions. Thus, it facilitates pre-training the mannequin on the baseline PV with the intention to make the most of the saved weights of the pre-trained mannequin and apply TL on the goal PV.”
The stacked LSTM mannequin considers temperature, humidity, photo voltaic irradiance, PV manufacturing, one-hot encoding illustration of the month of the 12 months, and sine/cosine transformation of the hour of day. The three methods had been based mostly on three completely different approaches: protecting the weights of the layer mounted, fine-tuning the weights of the layer based mostly on the goal area information, and coaching the weights of the layer from scratch based mostly on the goal area information.
The “TL Technique 1” method is reportedly in a position to extract options from the supply area and carry them to the goal area.
“This can be a extensively used scheme when treating photographs, the place the primary layers are used as function extraction layers and the final layers are used to adapt to new information,” the researchers defined.
With “TL Technique 2,” weights of all layers of the TL mannequin are initialized based mostly on information from the supply area. They’re fine-tuned based mostly on information from the goal area.
“This method is extensively used with issues the place there may be an abundance of information within the supply area, however a shortage of information within the goal area,” the group stated.
In “TL Technique 3,” the preliminary layers of the TL mannequin are frozen and the final layer is educated from scratch, popping the final layer of the bottom mannequin and including a brand new layer after the frozen layers.
“This method is much like the primary one, nevertheless it differs in the truth that the weights of the final layer are usually not initialized based mostly on information from the supply area,” the teachers stated.
The researchers used the three methods to forecasting the hourly manufacturing of six photo voltaic crops situated throughout a number of areas in Portugal. Their effectiveness was in comparison with that of standard non-TL fashions.
“The findings of the experimental software point out that every one three TL methods considerably outperform the non-TL method when it comes to forecasting accuracy, evaluated by a number of error indexes,” the scientists stated. “Outcomes point out that TL fashions considerably outperform the traditional one, reaching 12.6% accuracy enchancment when it comes to root-mean-square error (RMSE) and 16.3% when it comes to forecast talent index with one 12 months of coaching information.”
They launched the mannequin in “Transfer learning strategies for solar power forecasting under data scarcity,” which was just lately revealed in Scientific Reviews.
“This examine is step one in direction of enhancing our understanding of the influence of TL on photo voltaic plant energy prediction,” they concluded. “Future work will focus on assessing the influence of the bottom mannequin’s coaching information quantity, investigating whether or not coaching base fashions with extra information or with information from completely different photo voltaic crops may additional enhance forecasting accuracy.”
