Python Fastai library - Loss and Validation interpretation

The following is a regression problem, where I have a dataset with composed by more than 40k consumption ids with some variables such as DateTime, month, year, temperature, humidity.
I built a Fastai Tabular Data format using Embedding Layers for categorical variables. The historical consumption values (my target) are in the range [0, 1.400.000].
I used the tabular_learner with two dense layers [1000, 500].

After creating the learner, I had the following result using lr.find()

So I used to fit my model using the learning rate equal to 1e-02.

learn.fit_one_cycle(3, max_lr=slice(1e-02), wd=0.2)

Unfortunately, I reached the following results with a RMSE which is increasing during epoches:

I think I have a multicollinearity problem, that’s why R^2 score is -inf.
Another consideration: why validation and train losses are so high?