tensorflow - LSTM 平稳时间序列预测不稳定模型

Question

我是 Tensorflow 2 的初学者，我正在尝试进行多时间序列预测，以使用 10 个时间步长的向量（接下来的 10 天，最后一个密集层 = 10）来预测铜（LME）的价格。我的预测结果不是很好，所以我读到通过使数据静止，模型效率更好。我修改了我的程序如下：

-Time serie stationary 
-2 Bidirectional layers(LSTM) of 20 neurons with 10 neurons in the output layer 
-SELU activation
-initialization lecun_normal
-lr : 0.0005
-batch normalization
-lookback : 45 + Batch_size :8 + mse + seed constant 4000
-Validation split: 0.05 + Dropout :0.01

序列平稳性较好，但模型不稳定。我认为我在选择超参数时犯了初学者的错误。的确：

- There is no correlation between the minimum loss and the prediction quality
- The result is totally variable with the modification of the seed value
- With a training with a larger number of epochs, the prediction curves flatten
- The smallest val_loss does not generate the best prediction
- Val_Acc does not exceed 0.14
- During training the loss slowly decreases but the val_acc does not increase.
- Often the best results are for less than 5 epochs of learning 
- Loss and val_loss decrease but val_accurate does not increase
- The model prediction is very sensitive to the parameters. You can compare the two predictions of `enter code here`the same model in dotted red over 10 days with Dropout=0.01 and Dropout 0.02: they are opposite.

Do you have an idea to improve my model?
Thanks 

DROPOUT : 0.01  上升预测 在此处输入图像描述 DROPOUT : 0.02  自上而下预测 在此处输入图像描述

标签： tensorflowtime-serieslstmprediction