Dear Sir,

Please check

https://towardsdatascience.com/time-series-machine-learning-regression-framework-9ea33929009a

Myriam -

This does look like time series with seasonality, using regression, but I haven't looked a lot at time series. The time of day seems important which might be a challenge to model properly. Perhaps you could use categories. As for heteroscedasticity, it should exist based solely on the fact that one prediction is larger than another. (For an explanation of heteroscedasticity under finite population sampling, for regressions of format y = y* + e, see https://www.researchgate.net/publication/320853387_Essential_Heteroscedasticity.)

I don't suggest any 'tests.' They are overused, often ambiguous, and don't tell you what to do.

I think you will have to see how your software handles things like autocorrelation, regression weights for heteroscedasticity, and what your options are for dealing with seasonality and preparing data. As for nonnormality of estimated residuals, this is generally not fatal.

As for ridge regression, did you use that because you have collinearity? Perhaps you have too many variables?

It may be useful to consider both a "graphical residual analysis" to check model fit, and "cross-validation" to study whether you are overfitting to your particular sample.

Hopefully someone better at time series can give you some areas to look into there, as my experience is much more in the area of finite population "prediction" (where prediction means 'estimation' for a random variable, not forecasting).

It looks like you have a rather complex problem, and need to explore your options. Whatever you do, if you then take out data you know, and see what you would substitute for it with your regression, that might at least give you some idea as to whether or not you have done something reasonable.

Cheers - Jim

PS - The link that Ajit gave you looks helpful, I think.