使用 Keras 进行长期时间序列预测，预测依赖于未来输入

Question

I am building a neural net with Keras for predicting the voltage output of a complicated electrical circuit based on known current injections at distinct input nodes. As of right now, the model is a simple ConvLSTM model, which produces fairly good predictions. The problem is that I need to predict several minutes of voltage output on a millisecond timescale based on the initial 50 ms of ground truth voltage and the known current injections, so as I move forward in time, the model starts to rely on it's own predictions, therefore the prediction error compounds. I can measure the behavior of the system to obtain training datasets, but it's really slow, hence the need for a NN.

In my current mdoel I can only predict the next voltage value in time as the current input changes every millisecond, therefore long-term forecasting is not feasible. What I want to do is to build a model which takes 50 ms of ground truth training voltage, 60 ms of current input (starting at the same time as the voltage values, but exceeding it by 10 ms), and 10 milliseconds of voltage as a target. This would potentially help me to strenghten the first prediction (51st datapoint).

Unfortunately no matter how long I searched, I haven't found an architecture which was capable of handling this kind of input. Does anyone have any idea for a basic keras architecture?

Thanks!

Answer 1

Viktor, I suggest to feed LSTM for training with combination of actual current (at t between 11 ms and 50 ms) and voltage from 10 ms ago (from 1 ms to 40 ms) then you have a rectangular input (2 x 40) for each training sample.