# [Metadata-Augmented Neural Networks for Cross-Location Solar Irradiation Prediction from Satellite Images](https://milets19.github.io/papers/milets19_paper_1.pdf)

###### tags: `references`

[Web Service](https://solarmax.thingnario.com:420/?email=[e-mail]&password=admin)

## Introduction
1. PR(Performance Ratio): the ratio between actual and theoretical energy ouputs
2. In the past, pyranometers are used for monitoring solar irradiance, but they are expensive and they need regular mentainance.
3. Cloud is decisive.
4. Nearby weather stations are not reliable.

## Datasets
1. Satellite irradiation dataset (SATI)
* SATI-Taiwan: 30 stations (3 for validation), 2015/7/7~2018/3/31
* SATI-BSRN: Fukuoka and Kwajalein, 2017/1/1~2017/12/31
* 1 satellite image every 10 min, 6 image a patch
* RGB visible channel: 550x550 pixels, 1km/pixels
* Infrared channel: 550x550 pixels, 2km/pixels
* Normalization: substract 127 and divided by 127 => pixels $\in (-1,1)$
2. Elevation and azimuth given time, date and geolocation.
3. Ground truth: irradiation from stations where pyranometers are under regular cleaning and calibration
4. Data Preprocessing: cropped the surrounding 25x25 pixels for both the visible and infrared channels, which were stacked to form a 4-channel patch.
5. Data Augumentation

## Proposed Method
1. Irradiation Predictor:
* ResNet16 for satellite image extractor: change the first 7x7 convolution to 3x3, remove maxpooling
* LSTM for capturing the temporal relationship between the images

2. Metadata Encoding:
* Dimension of elevation one-hot: 360
* Dimension of azimuth one-hot: 180
* The one-hot vectors are transformed into an attention map and then multiplied on the satellite image to be the final input of ResNet.

3. Spatiotemporal Recurrance Autoencoder
* Reconstruct the cloudy and cloudless image
* First feature: reconstruct cloudless image. Capture the surrounding landscape information.
* Second feature: Capture cloud information
* Training loss: MSE

## Experiments
1. Evaluation Metrics:
* rRMSE(Relative Root Mean Square Error): The spread of errors. Compare across seasons and station.
* MAE: The deviation from ground truth.
2. Metadata Fusion:
3. Results:
4. Ablation study: