I have implemented a neural network for load forecasting in Microsoft Excel. My structure is very simplistic and involves only 1 hidden layer and 4 neurons. (See picture)

Neural Network Structure

I trained my network with a very large set of data and the predictions are as decent as could be expected. However, when I go to try and predict the "future" using recent data, the network is only in the ball-park, not very close at all. Here is a plot of the actual numbers versus the forecast ones.

Neural Network Prediction Error

From looking at this chart, can it be determined if I need another layer or more neurons??


There isn't any hard and fast rule to determine that how much hidden layers would work. In your case, extra layer won't help that much, but extra features can (I'm pretty much sure). Weather forecasting, as far as I know, isn't something which can be determined using only 4 features (considering your input neurons). You can always come up with more and more features, which effectively help in predicting the weather. So, I'd suggest to look for more features (about 10 to 15 useful features), and come up with the following configuration of the network:
1.) Input: Number of neurons in he input layer should be equal to the number of features you calculate for each data entry.
2.) Output: If you can categorize the weather in, say, 5 different categories (sunny, rainy, foggy, stormy, meteor shower), then there can be 5 output neurons.
3.) Hidden: Number of neurons in the hidden layer is normally the mean value of the number of input, and output neurons. You can also have same number of neurons in the input, and hidden layer.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.