ANFIS EDIT out of memory

I am currently uploading some training data, which includes 10 inputs and 1 output. The data consists of the following variables: day, time, water heater temperature, water heater flow rate, and the frequency of switching on and off hourly. The output variable is the predicted flow rate for the next week. In total, I have data for 4 weeks and corresponding predicted flow rates.

However, when running the ANFIS, I am encountering an 'out of memory' error. I am also uploading the data here what could lead to this. Perhaps I am missing something.

Thank you.

[Walter Roberson]

It would help to have your code.

[Antonio Abela]

I am using the ANFISEDIT tool. Is there any possibility that I can code it?

Answer 1

[Sam Chak]

Hi @Antonio Abela

You mentioned that there are 10 inputs and 1 output. However, in the CSV file, there are 15 columns, with each column containing 168 rows. It seems that you may have selected 11 columns out of the total.

If you utilize the grid partitioning method and uniformly assign 5 membership functions {VeryLow, Low, Medium, High, VeryHigh} to each input, you will end up with approximately 10 million rules. Consequently, when the ANFIS attempts to undergo training, the extensive data allocation can lead to "out-of-memory" issues.

num_MFs   = 5;
num_Input = 10;
num_Rules = num_MFs^num_Input

num_Rules = 9765625

[Antonio Abela]

Yes, sorry for the confusion. I mistakenly swapped the values. In total, I have 14 inputs. As previously mentioned, these include the day of the week, the hour, and the state (whether it is on or off) during that hour, as well as the temperature of the water tank and the flow rate. These values are repeated four times, representing four weeks.

Could you please suggest any alternative approaches I could use for this? I am open to suggestions.

[Sam Chak]

Hi @Antonio Abela

Let me ask you a question. Are the fuzzy rules of utmost importance to you? Or do you simply want accurate predictions without the need to comprehend the underlying correlational reasoning (important in Social Science) provided by the fuzzy system?

If the latter is the case, I would suggest training your data using deep learning neural networks. With this approach, you can obtain predictions without the necessity of understanding and interpreting the outputs generated by the fuzzy system model. Just imagine dealing with billions (nine zeros) of rules for analysis and interpretation!

num_MFs   = 5;
num_Input = 14;
num_Rules = num_MFs^num_Input

num_Rules = 6.1035e+09

[Antonio Abela]

Hello @Sam Chak,

No, I just want to have an accurate prediction; the rules are not very important. Could you kindly provide me with more insight on this? If possible, could you share some examples or explain how I can approach this? Your assistance would be greatly appreciated.

[Sam Chak]

Hi @Antonio Abela

Since you have the data points collected over time, you can apply a deep learning-based time series forecasting technique by using past observations of time-dependent variables to make predictions about future values of the desired variable.

I recommend reading this article on time series forecasting using deep learning:

https://www.mathworks.com/help/deeplearning/ug/time-series-forecasting-using-deep-learning.html.

It provides detailed explanations and examples to help you understand the concept. Once you have gone through the material, feel free to share your code and discuss it here.