Add model

README.md

@@ -0,0 +1,49 @@
+---
+license: mit
+tags:
+- nowcasting
+- forecasting
+- timeseries
+- remote-sensing
+---
+
+# Nowcasting CNN
+
+## Model description
+
+3d conv model, that takes in different data streams
+
+        architecture is roughly
+        1. satellite image time series goes into many 3d convolution layers.
+        2. nwp time series goes into many 3d convolution layers.
+        3. Final convolutional layer goes to full connected layer. This is joined by
+        other data inputs like
+        - pv yield
+        - time variables
+        Then there ~4 fully connected layers which end up forecasting the
+        pv yield / gsp into the future
+
+## Intended uses & limitations
+
+Forecasting short term PV power for different regions and nationally in the UK
+
+## How to use
+
+[More information needed]
+
+## Limitations and bias
+
+[More information needed]
+
+## Training data
+
+Training data is EUMETSAT RSS imagery over the UK, on-the-ground PV data, and NWP predictions.
+
+## Training procedure
+
+[More information needed]
+
+## Evaluation results
+
+[More information needed]
+

config.json

@@ -0,0 +1 @@
+{"include_pv_or_gsp_yield_history": true, "include_nwp": true, "forecast_minutes": 120, "history_minutes": 30, "number_of_conv3d_layers": 6, "conv3d_channels": 32, "image_size_pixels": 24, "nwp_image_size_pixels": 64, "number_sat_channels": 11, "number_nwp_channels": 1, "fc1_output_features": 128, "fc2_output_features": 128, "fc3_output_features": 64, "output_variable": "gsp_yield", "embedding_dem": 8, "include_pv_yield_history": true, "include_future_satellite": false, "live_satellite_images": true, "gsp_forecast_minutes": 480, "gsp_history_minutes": 120, "include_sun": true}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02aed3680db164d9ece24b158d5facaf34fc03687bb36d3ed19362acf1a86cb3
+size 181497514