- the data is from link
- first of all, I train a model MSTAN, the paper is Multi-Source and Temporal Attention Network for Probabilistic Wind Power Prediction, but the result is not good at all..
- then we trained the Seq2seq mdoel, very simple but it works very well..
- See the result at show_data.ipynb
- 15min a point, use 48 data to predict the next 48 data
- the feature that we use is just wind speed at 10 meters, wind direction at 10 meters , The rest of the data is not very impactful, but adding them will improve performance a bit
- The result is quite good with a MSEr about 3.2 for a turbine capacity of 100MW
- 对于概率预测，我觉得直接去预测概率模型的分布，是是不合理而且效果很不好的。最一开始我是用MSTAN去预测一个Beta分布的概率，最终结果非常不好
- 一个比较好的概率预测可以这样的到：使用多个模型预测同样的结果得到一个序列[x1,x2,x3,x4...xn]，然后对这个序列进行KDE(kernel density estimation)分析，得到的概率结果可能才会更可信一些。