Question about 'nan' value in the evaluation process

[LiuZhihhxx]

I just used the baseline code that officially released to train the model and submitted my codes. Howeveer it yielded "NaN in predicted values" error. I tried to debug locally step by step only to found out an interesting bug. There are 134 models correlated to 134 turbines respectively, and in evaluation some models worked well and generated valid predicitons while else generated ALL 'nan' value. As all models and test data have the same structure for each turbine, it seems this bug should not happen.

Anyone have similar problems? Any help plz~