This markdown helps to understand the codes and the folder(file) structure.

 Team CSU123:
  test_x.zip, test_y.zip: test data for evaluation

 \paddlepaddle:
  common.py, metrics.py, model.py, predict.py, train.py, test_data.py, wind_turbine_data.py
  are almost same as officially released.
  data_cleaning.py: KNN to deal with null and abnormal data,

 \paddlepaddle\sample:(partial hyperparameter grid search)
  Similar structure to \paddlepaddle. eva luation.py and train.py were modified.

 1.\paddlepaddle\sample\prepare.py
  A parameterized prep_env_search() is provided based on the original prep_env().It enables variable hyperparameters(batch size,learning rate, number of layers, and input length).
 2.\paddlepaddle\sample\train.py
  A multi-stage cycle is designed to realize grid search. That is, train models with different hyperparameter combinations. The trained models were saved at \paddlepaddle\sample with seperate folders.
 3.\paddlepaddle\sample\evaluation.py
  MAX_RUN_NUM was changed as 1 to avoid error.
 4.\paddlepaddle\data_cleaning.py
  A method is defined to realize kNN interpolation by kNNImputer.
 5.\paddlepaddle\make_zip.py
  Compress all files into a zip file according to the submission format requirements.

 1.Run \paddlepaddle\preprocess.py and get file 'wtdata_245days_knned.CSV'.
 2.Run \paddlepaddle\sample\train.py to train models on sampled turbines(can be defined in \sample\prepare.py).This yields a group of foldersin \sample\checkpoints where models exist.
 3.Run \paddlepaddle\make_zip.py. This transform the model folders into .zip files so that satisfy the evaluation procedure.
 4.Run \paddlepaddle\sample\evaluation.py to evaluate different models.(For some unresolved reason about temporary system path, the eval method can only launch once in a single run, otherwise 'path not found error' will occur. In practice, we record every score manually.)
 5.Run \paddlepaddle\train.py to train 134 models using the best set of hyperparameters in step 4(by set prepare.py).