2:03 PM - 2:18 PM
[MZZ48-02] A study on information compression of meteorological image using Auto-Encoder for area photovoltaic power prediction
Keywords:photovoltaics, area photovoltaic power prediction, Auto-Encoder, Convolutional Neural Network, latent space
The output of photovoltaics (PV), which are recently much penetrated to power systems, intrinsically fluctuates depending on the solar irradiance. In power systems, the demand-supply adjustment of electric power is indispensable. Therefore, the energy managements with the prediction of PV power generation attract much attention. In the case of large prediction error, it can be extremely difficult to adjust the supply and demand of the power system. So, the PV output prediction is required to have the ability to reduce a large error as well as an average error. Authors have carried out the prediction of area PV power in the Tokyo Electric Power Company in Japan [1]. We developed the novel prediction method of area PV power by means of dimensionally compressed information of regional meteorological image by the Auto-Encoder [1]. In this paper, we investigate the appropriate number of dimensions of the latent space in the Auto-Encoder focusing on the case that the three kinds of meteorological images are input to the developed prediction model, i.e., global horizontal irradiance (GHI), low cloud amount and accumulated rainfall.
Fig. 1 shows the developed prediction model composed of the Auto-Encoder using convolutional neural network (CNN)[1]. The Auto-Encoder is one of the neural network structures, which has low dimension latent space inside of networks. We perform training of networks so as to reproduce the input meteorological image through Decoder from low dimensional data in latent space compressed by Encoder, which results in the network that can dimensionally compress the input image information while properly retaining the characteristics of the image. In the developed method, after compressing the meteorological image information, fully connected neural network, which is trained at the same time as the Auto-Encoder, predicts area PV power using the compressed data in latent space. The developed method enables us to perform precise prediction of area PV power utilizing huge amount of meteorologically spatial information with lower computational load.
The meteorological images are made from numerical data by the Meso-Scale Model (MSM) of Japan Meteorological Agency (JMA) which is member 00 (control run) of the Meso-scale Ensemble Prediction System (MEPS). We arrange 60 x 60 pixel images by using the MSM prediction value calculated at intervals of 5km in the Kanto area of Japan (around the Tokyo in Japan) for the three kinds of meteorological element such as GHI, low cloud amount, and accumulated rainfall [1].
Under the condition explained above, we investigate the appropriate number of dimensions of the latent space by carrying out the prediction for 4 months while changing the dimensional number. As a result, it is revealed that approximately 15 dimensional latent space is enough for the compression of 60 x 60 x 3 size input information from the viewpoint of prediction accuracy.
In the future works, we will improve the prediction accuracy by inputting the information of the MEPS members, i.e., ensemble run, in addition to member 00 (MSM) into developed prediction model.
Acknowledgment
We would like to thank the Japan Meteorological Agency (JMA) headquarters and Meteorological Research Institute (MRI) of the JMA for cooperation of using MEPS data.
References
[1]Y. Mori, S. Wakao, H. Ohtake, T. Takamatsu, T. Oozeki, “A fundamental study on PV output forecast over wide area by means of Auto-Encoder,” IEEJ, The papers of Technical Meeting on Frontier Technology and Engineering(2021-12-1), pp.67-72, 2021 (in Japanese)
Fig. 1 shows the developed prediction model composed of the Auto-Encoder using convolutional neural network (CNN)[1]. The Auto-Encoder is one of the neural network structures, which has low dimension latent space inside of networks. We perform training of networks so as to reproduce the input meteorological image through Decoder from low dimensional data in latent space compressed by Encoder, which results in the network that can dimensionally compress the input image information while properly retaining the characteristics of the image. In the developed method, after compressing the meteorological image information, fully connected neural network, which is trained at the same time as the Auto-Encoder, predicts area PV power using the compressed data in latent space. The developed method enables us to perform precise prediction of area PV power utilizing huge amount of meteorologically spatial information with lower computational load.
The meteorological images are made from numerical data by the Meso-Scale Model (MSM) of Japan Meteorological Agency (JMA) which is member 00 (control run) of the Meso-scale Ensemble Prediction System (MEPS). We arrange 60 x 60 pixel images by using the MSM prediction value calculated at intervals of 5km in the Kanto area of Japan (around the Tokyo in Japan) for the three kinds of meteorological element such as GHI, low cloud amount, and accumulated rainfall [1].
Under the condition explained above, we investigate the appropriate number of dimensions of the latent space by carrying out the prediction for 4 months while changing the dimensional number. As a result, it is revealed that approximately 15 dimensional latent space is enough for the compression of 60 x 60 x 3 size input information from the viewpoint of prediction accuracy.
In the future works, we will improve the prediction accuracy by inputting the information of the MEPS members, i.e., ensemble run, in addition to member 00 (MSM) into developed prediction model.
Acknowledgment
We would like to thank the Japan Meteorological Agency (JMA) headquarters and Meteorological Research Institute (MRI) of the JMA for cooperation of using MEPS data.
References
[1]Y. Mori, S. Wakao, H. Ohtake, T. Takamatsu, T. Oozeki, “A fundamental study on PV output forecast over wide area by means of Auto-Encoder,” IEEJ, The papers of Technical Meeting on Frontier Technology and Engineering(2021-12-1), pp.67-72, 2021 (in Japanese)