Seeking Advice: Creating Tide Forecasting model

I am expected to build a tide forecast model at my internship. The next step after the model is ready would be to have it work on real time data. I have about 3 months of data (with 2 columns - Timestamp and Tide Value) with a very small gap at the end of the 3rd month as of now (I have asked for more data and should be receiving it soon).

So far here is what I have done:

  • I applied FFT to the data and then reconstructed to get the frequency, amplitude and phase of the data in a csv
  • I was unable to use the frequency domain data in the prediction model though cause I am still trying to figure out how
  • I used to the darts library to fill in the small part of the data and then used its FFT Forecast Model to get the forecast
  • The training and validation split was 90/10
  • The forecast I received started to fall off at the end
  • The MAE was 0.35090170672342885
  • Following the prediction plot: a9be0b106371846e544dd6a981b23910.png - Google Drive

I need help moving as I am new to the field.
Are there other approaches I can take?
Should I be using another model for this dataset?
Can I use the frequency domain data to make the model better?

Time series data is usually handled using a Recurrent Neural Network.

Hi Shivam. As you are new to working with time series data I suggest doing this Time Series course on Kaggle Learn:
Learn Time Series Tutorials | Kaggle

I also suggest reviewing the notebooks relating to the “Store Sales - Time Series Forecasting” Kaggle competition. They will give you some ideas for different approaches to use to solve your problem. You can start by reviewing the notebooks with the most upvotes.

One of the things you’il learn from the above resources is the importance of creating the train/val split correctly. You mention 90/10, but it’s important to ensure that train data is from the past and val data is from the future.

1 Like