HELP, how can I improve the neural network?

Hi everyone, I really need help from someone who knows the subject and knows how to address me correctly. I am creating neural networks for the prediction of the reactive power at the nodes of the transmission network, having as input to the neural network the active power and indication of month, day, hour, day of the week and type of day.

Training the network on the data of 2017 and 2018, obtaining not excellent but still decent results. The problem arises when I go to use the net for the 2019 reactive forecast. The net fires random values ​​and has sudden high peaks. The data I use is quarterly, I have an input file of 70080 lines for training; i am using a non linear input-output network to make the prediction. Can anyone give me directions on how to improve network performance? The autocorrelation and input-error cross correlation errors are very high, I tried to increase the input delay but by doing so I can only use the trainscg which, on the 2017 2018 data gives me good results but then on the 2019 data is completely lost. Is the network overfitting? Please I really need it. I am attaching an example of the script that I am using and the graphs I get for 2017/2018 and 2019.

An infinite thanks to those who will answer me.

[KSSV]

It is better to attach data as well. People who can answer can give a try to tun the code.

[Giuseppe D'Amico]

unfortunately I cannot attach the data, because they are sensitive data.

Answer 1

[Krishna]

Hello Giuseppe,

Based on your query, it seems you're working on seq2seq modeling with the ‘timedelaynet’, where active power is the input for predicting reactive power. To enhance your results, you might consider a few adjustments:

You're currently using ‘timedelaynet’, which only incorporates input delays. A viable alternative could be ‘narxnet’ in MATLAB, which includes both input and feedback delays. This adjustment could improve predictions by leveraging previous outputs to forecast future values, a critical aspect for some problems. To learn more about ‘timedelaynet’ and ‘narxnet’ please go through the following documentation:

https://www.mathworks.com/help/deeplearning/ref/narxnet.html

https://www.mathworks.com/help/deeplearning/ref/timedelaynet.html

Also, instead of solely focusing on autoregressive networks, consider exploring RNNs and LSTMs. These networks are highly optimized for seq2seq modeling challenges. Please go through the following documentation to know how to implement them in MATLAB,

https://www.mathworks.com/help/deeplearning/ug/long-short-term-memory-networks.html

Also, please ensure you include a comprehensive set of features that accurately represent the transmission network. A scarcity of features could lead to unpredictable network behaviour.

If there's a concern that your network is overfitting the test data, it's crucial to employ proper validation and testing datasets. If overfitting is indeed an issue:

• Consider using regularization techniques.
• Experiment with simplifying the network architecture to prevent overfitting.

Regarding the spikes observed in the 2019 data, it's challenging to pinpoint the exact cause without examining the dataset beforehand. Ensure that the data you're using is reliable and free from noise.

Hope this helps.