Temperature Prediction Using Recurrent Neural Network

A reshaping of vectors from 2D to 3D is needed since the the recurrent net wants to know the time frame its working from, in this case 24 hours.I decided on a 3 convolutional layer followed by 3 GRU units then followed by one dense layer for my final neural network architecture..GRU, a time sensitive architecture, uses an update gate and reset gate..Two vectors which decided what info should be outputed..They can be trained to keep information from long ago, without diluting it through time..The update gate decides how much of past to retain, the reset gate decides how much of it to forget.GRU cellThis architecture is peppered with dropout of 0.25 throughout, as I found otherwise it quickly overfitted due to the amount of details it captures..The activation function for the final dense(classification) layer has to be the “linear” function, since this is essentially sophisticated linear regression.After compiling and running the model, this is how the model improved over time:Loss over timeIt converges to about 0.02 relatively quickly..Overall a rather successful experiement.Is trying to predict the future ever a good idea?.It seems that without divine understanding of the golden ratio, or other such supernatural insight, prediction is never 100%, especially for a man made system such as the stock market, although Max would probably disagree..Thankfully, nature is a little more consistent..For practical purposes, this model would be great for catching anomalies in the weather pattern, for catching black swan events such as extreme weather events.. More details

Leave a Reply