Dear scholars,
Imagine there is a non-linear system, Like PCM storage tank (or ice storage tank, ... ). For optimization purpose or real-time control (reducing computational time) I need to make a surrogate model that can predict the performance of the real system until the future horizon, like next 24 hours.
Systems like thermal storages (imagine melting of ice if you're not familiar with the case) have large amount of inertia. During their melting or solidification, which takes a lot of time in my case, we have no data that what's going on in storage tank (temperature is constant). We just know how much time is passed.
I want to use ANN (artificial neural network) or Deep learning to make the black box of the mentioned system. But first, I need to be sure about the accuracy and performance of these methods for my purpose.
I would appreciate any help or advice.
Thanks.
Dear Mehrdad Vojdani,
I suggest you to see links and attached files on topic.
https://www.soderbergpartners.se/globalassets/sv/om-oss/karriar/arets-finansuppsats/bjorklund-s.--uhlin-t.-artificial-neural-networks-for-financial-time-series-prediction-and-portfolio-optimization.pdf
https://pdfs.semanticscholar.org/c6b8/f36c781a2186e331bba064be0ed5d109098f.pdf
https://www.kdd.org/kdd2019/proceedings
Conference Paper Comparative Study of Short-Term Wind Speed Forecasting Techn...
Hi Mehrdad Vojdani,
First of all you need to make sure that the amount of your data samples is quite enough to represent the system behavior. Based on my experience, the important aspect is the sampling period of your data. Since you mentioned about prediction in the next 24 hours and the system is a process control, I expect that the transient time is slow and your sampling time could be slower also, so most probably you should collect the samples for a longer period (2-3, days or even a week, it depends on your machine characteristics actually).
The next step, make sure your NN or Deep NN model complexity is match with your problem situation, to achieve your desired accuracy. You are able to design your own structure (let say: 5 layer and 5 nodes for each layer) and test with the data samples that you have, observe the results and try to adjust by increasing or decreasing the number of nodes and layers manually. Of course, this will take huge efforts and seems to be wasting time and you might come up with over-fitting or under-fitting condition or bias and variance trade-off dilemma. (I have one python code in jupyter example for a simple NN prediction, you can find here: https://github.com/abdulhady89/NNprediction)
There are several approach to find the best/optimal NN model. In our research group, we are also dealing with that problem on data stream concept view. Recently, my supervisor (Mahardhika Pratama) and teams developed an algorithm so called Neural Network with Dynamically Evolved Capacity (NADINE) which has auto-construction of network capability, you can take a look here: Conference Paper Automatic Construction of Multi-layer Perceptron Network fro...
Regards,
Abdul Hady
hello
you can use simulation data for leaning and test for example you can use Matlab or any software and model it by mathematical equations and use it. it helps you with the validation result too.
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling