convolutional neural network for regression

I have a [17000 obs,72 statistical features] dataset of sky images, and I want to train a CNN for regression problem (sky image --> irradiance) but I don't know how. Since I don't have statistical features as input rather than pixels of images, all solutions in the documentation do not help. Is there any other solution I can try? FYI: I've only turned to deep learning because standard neural networks were giving me a bit high error values.. Thank you for your time

[Greg Heath]

Have you plotted the irradiance vs each of the 72 inputs to try to get a feel for weak and strong I/O dependencies?

Are there distinct regions where an output's dependencies on the input seem to be quite different from other regions others?

Have you investigated the interdependencies of the inputs? I doubt that you need all 72.

Hope this helps.

Greg