Fine tuning an LSTM Encoder-Decoder model for multi-step forecasting
Budget: €30 – €250 EUR
I built an LSTM Encoder-Decoder model for multi-step forecasting, it is many to many, one feature only. it includes the timestamp and its value.
I used a sequential model with input, hidden, and output layers. I tried adding more hidden layers but it did not produce a good accuracy, If I increase the number of neurons, then the model gets nan in the training step, therefore, I'm using the model on lower neurons. I feed 240 timesteps in, and forecast 240 timesteps out (5 days) by single-shot predictions.
I need a professional in this field who can find out what is the problem in my implementation, so I can improve the forecasting accuracy in the long run.
The tasks of this job:
- Find out the logical error of the model, and fix it.
- Documentation/comments on each modification step.
- Error (MAE) allowed until 10% of the maximum value of the actual data.
you can find the script (implementation), and the data in the attached files.
I used a sequential model with input, hidden, and output layers. I tried adding more hidden layers but it did not produce a good accuracy, If I increase the number of neurons, then the model gets nan in the training step, therefore, I'm using the model on lower neurons. I feed 240 timesteps in, and forecast 240 timesteps out (5 days) by single-shot predictions.
I need a professional in this field who can find out what is the problem in my implementation, so I can improve the forecasting accuracy in the long run.
The tasks of this job:
- Find out the logical error of the model, and fix it.
- Documentation/comments on each modification step.
- Error (MAE) allowed until 10% of the maximum value of the actual data.
you can find the script (implementation), and the data in the attached files.