Question:

I am working on a predictive model for cumulative energy consumption of EV based on short-term time series data. The dataset consists of approximately 1000 samples, where each sample includes 30 time steps of sensor data, with 4 features at each time step. The target value is the cumulative energy consumption at the final time step.

Approach so far:

I have experimented with both CNN-based (ResNet) and RNN-based (LSTM) architectures.
However, the prediction accuracy has not been satisfactory with either of these methods.
Problem:

I aim to optimize the prediction specifically for the cumulative energy consumption at the last time step.
I’m concerned that simply optimizing for the last time step might lead to issues or not fully utilize the sequential nature of the data.

Questions:

Architecture: Are there any specific architectural adjustments I should consider for this task? For instance, would a combination of CNN and RNN architectures (e.g., ConvLSTM) be more effective?
Data: Given the short sequence length (30 time steps), are there any data preprocessing or augmentation techniques that could help improve model performance?
Loss Function: Should I consider using a custom loss function that places more emphasis on the accuracy of the final time step, or is there an alternative loss function better suited for this type of prediction?
Other Considerations: Are there any other modeling techniques, regularization methods, or optimization strategies that might be particularly useful for improving the accuracy of cumulative predictions in short-term time series data?
Any insights or suggestions would be greatly appreciated!