Description
Description:
Reading the section 10 Deep learning for timeseries, there appears to be a potential discrepancy between the problem statement and the actual implementation.
Problem Statement:
The problem statement, as described in section 10.2.1, outlines a scenario where temperature data and other variables for 5 days, sampled once per hour, are provided. The objective is to predict the temperature 24 hours ahead.
Concern:
According to the problem statement, there are 120 samples in 5 days (24 samples per day). The dataset should consist of sequences representing 5 days of data, with each sequence containing a maximum of 120 samples.
Keras Implementation:
However, when utilizing the timeseries_dataset_from_array() function with parameters sampling_rate = 6 and sequence_length = 120, it generates sequences corresponding to 30 days (4 samples per day). This seems to deviate from the problem statement's objective of predicting temperature with data from 5 days, not 30.
Proposed Solution:
One potential solution could be adjusting the sequence_length parameter to 20. This adjustment would ensure that sequences contain data from 5 consecutive days (4 samples per day using sampling_rate = 6), aligning with the problem statement's requirements.
Request for Clarification:
I'd appreciate clarification on whether my analysis is accurate and if the implementation aligns with the intended problem statement. If not, guidance on how to correctly utilize the timeseries_dataset_from_array() function for the specified problem would be valuable.
Thank you for your attention to this matter.