What model should be used for vector to scalar time series prediction

Question

I'm new to machine learning and I'm trying to work on a problem and the data looks like the following:

Date/Time            Tchg   Tcold   Pchg    Qchg    Flow    CUF  dCUF
1999-07-20 00:00:00 479.606 550.187 1238.22 94.1212 0.0 0.00492  0.00001
1999-07-21 10:20:00 480.169 550.187 2238.22 94.3004 0.0 0.00492  0
1999-07-21 00:00:00 479.606 550.187 1238.22 94.1212 0.0 0.00496  0.00004
1999-07-22 00:17:42 279.606 550.187 1238.22 94.1212 0.0 0.00496  0
1999-07-22 03:43:03 279.606 550.187 1238.22 94.1212 1.0 0.00496  0
1999-07-22 00:00:00 279.606 350.187 2238.22 94.1212 1.0 0.00497 10.000011

1. The data set has 6 columns with 5 of them will be input and CUF to be predicted. The dCUF is just the change of CUF from previous row.
2. During each day the 5 inputs will be collected from sensor for several times and each time they may change. Based on the each collected data, CUF may change but it's only updated at the end of the day.
3. Each row input actually may create a change of CUF from previous row (time history) but it's not shown.

I'm thinking of combining the input for each day which is an array of n x 5 (n depends on how many times the input data are collected and is not fixed) and then try to predict the dCUF which is the change of CUF for each day.

My questions are:

1. Will this be predictable since we don't have correct CUF at each collected time?
2. If this is predictable, what kind of data manipulation should I do and what model should I use?

Answer 1

I assume the 5 columns are parameters for CUF then there should be a mathematical formula or model for CUF. Anyway since there is a variation in data you need Time-Series data prediction and I think the only way to do it is to build neural network model that you can train and has five inputs (Tchg,Tcold,Pchg,Qchg..) and CUF as output layer. Tensorflow and pytorch are all good starting point for building CNN check here:

https://www.tensorflow.org/api_docs/python/tf/keras/layers/Conv1D

and also hybrid approach (quasi-recurrent neural networks, QRNN) as discussed by Bradbury et al (2016) https://arxiv.org/abs/1611.01576

Answer 2

1. If CUF is your objective variable, and it is not measured correctly, there is no way to have an ML model that predicts the right value, if the right value is not known.

2. Time series forecasting using multi-input variables is normally addressed using RNN models like LSTM or CNN, check this tensorflow tutorial.

As for the preprocessing, a good rule of thumb is batch normalization when dealing with neural networks, others techniques, highly depends on the problem.