Page 208 - AI for Good-Innovate for Impact Final Report 2024
P. 208

AI for Good-Innovate for Impact



                      maintenance due to its intermittent nature. Therefore, accurate nowcasting and day-ahead
                      forecasting of the green power is crucial in power dispatch and electricity trading. Traditional
                      methods usually depend on simple regression models and thus suffering low accuracy.
                      Meanwhile, the forecasts were always deemed as black-box outputs to grid operators thus
                      short of interpretability.
                      In this use case, we applied AI based methods to deliver accurate and interpretable renewable
                      energy forecasting in an Asian city.  We adopted convolutional neural networks and conventional
                      tree based models with large scale automatic feature augmentation to produce reliable wind
                      power and solar power forecasts. We used a shapely value based XAI (Explainable AI) method
                      to interpret why the model makes such forecast, via visualizing the contribution from different
                      weather factors.

                      The forecasting service covers all the wind plants and the rooftop photovoltaics within the
                      area. The interpretable analysis along with the forecasts to enable people understand how
                      the model works.

                      UN Goals:

                      •    SDG 7: Affordable and Clean Energy
                      •    SDG 9: Industry, Innovation, and Infrastructure
                      •    SDG 11: Sustainable Cities and Communities

                      Justification UN Goals selection: In this use case, we use AI based models to produce reliable
                      wind power and solar power forecasts, which entails the grid operators to make confident
                      decisions for power dispatchment and  real-time electricity balancing. It helps the digestion of
                      the economic green energy to the electric grids inspite of the intermittency challenges, which
                      contributes to SDG 7: Affordable and Clean Energy. In terms of SDG 9: Industry, Innovation,
                      and Infrastructure, our use case has been deployed in grid operation system for about 2 years,
                      which is widely regard as an innovative system in the power systems industry. Last but not least,
                      our work is committed to build a better city with more clean energy, which fits well with SDG 11:
                      Sustainable Cities and Communities.


                      48�2�2� Future work

                      Given more resources, we would like to elaborate the use case featuring the improvement of
                      forecasting performance under complex weather scenarios such as tropical cyclones, storm
                      and icing. To achieve this goal, we would need more data including the measured weather
                      conditions (e.g. sensor data, satellite images, etc.) at each power plant, historical complex
                      weather period, longer history of measured power, to enrich our model training and pay more
                      attention to those complex scenarios. More advanced machine learning techniques such as
                      Transformer based models may be applied to incorporate the various data modalities.


                      48�3� Use Case Requirments

                      •    REQ-01: It’s required to use numerical weather prediction data for day-ahead forecasting.
                      •    REQ-02: It’s required to use measured power data from renewable power plants, which
                           is always private, for model training.
                      •    REQ-03: It’s required to use real-time observed weather data for error analysis.
                      •    REQ-04: It’s recommended to use cloud based data transmission for inner deployment.




                  192
   203   204   205   206   207   208   209   210   211   212   213