用于具有多个输入的时间序列函数分类的Tensorflow模型

Question

I am new at Tensorflow.

I would like to classify the result of a multiple of (6 or more) timeseries floating point timeseries inputs into three predictive future output categories: large positive change (above a preset level), large negative change (below a preset level), not much change positive or negative (in between the preset level).

In the training and test sets, the multiple inputs are used to predict a single fp value in the future. Each of the inputs are fp historical values evaluated at regular time intervals. We are not interested in the exact fp output value in the future. We are only interested whether the change in the output value, between the current time and the future time, will be largely positive, largely negative or not much change (in between).

The training and test datasets will have the historical input values and the output classification labels (LP, LN, and BTWN).

Can anyone suggest a good tensorflow model for this application?

For example, let's say that you have a sin function advancing at a regular angular rate and you want to predict what the change in the sin function will be at +10 degrees (time) in the future. You can use just 1 fp input (noisy sin value) and you can predict the change of the sin value over the next (future) +10 degrees will be either LP, LN or BTWN some preset value. In this example, near 0 and 180 degrees of rotation the rate of change in sin value would be LP or LN and near 90 and 270 degrees the rate of change would be low or in between some preset threshold.

Now with 2 input functions, noisy sin and noisy cos, the system should be able to do a better prediction of the change in the sin value in the next +10 degrees.

If we add some additional related noisy input functions like atan, angle, and circumferential position, we should be able to get better and better predictions of the future change in the sin value with a high degree of certainty even if the noise levels are significant.

My application does not concern trigonometric functions, but in a similar way the input functions are all noisy but can be used to aid in the predicted classification of the future change in the output function as being Largely Positive, Largely Negative, or In Between and that will be sufficient to achieve the desired goal. Predicting the exact numerical value of the output function is not necessary, the classification into the 3 output categories will serve us well.

The output could be the result of only 2 binary digits where 10 = LP, 01 = LN and the other values are uncertain or In Between the preset threshold value.

So, can you suggest a starting tensorflow model for this application?