How to label an RBG image consisting of 8 bits/ data points and each point represents a certain power value to use the image as an input to CNN for forecasting

I have a set of power values recorded each hour, then created RGB images of those power values so each image represents 8 hours. The image attached represents 8 different power values (since each hour represents a single power value). I want to use those images as an input to CNN, but don't know how to because I can't label the image.

How can I label the images so I use them as an input to the CNN? Is there a code to do that on MATLAB?

[Pratham Shah]

Can you be more specific about the image? Like, what the X & Y co-ordinates represent in the image?

[Danya]

The X co-ordinate represents time while the Y co-ordinate represents power.

The image was created by converting power signal to an image. Each 1 hour has an alternative power value. The attached picture below contains 8 different colours and each colour represents a power value. I drew lines above the image to make it clear that it contains 8 different colours

[Pratham Shah]

Can you please state the problem statement?

I believe power value is represented by the pixel intensity not Y co-ordinates. Your problem statement will help me to understand the image. Becasue I can see some white pixels in the 3rd hour.

[Danya]

I agree with you that the colour intensity represents the power. Here is a detailed background explanation to make the question clear:

The power data was reconstructed into 2D image forms.Those images contain hourly input correlations of the power plant. Prior to the reconstruction process, the wind power data was normalized with a min–max normalization method in order to ensure that all features were distributed according to the same scale. The last 8 h were chosen for the image form. Therefore, each set of images consists of 2 × 4 array sequentially and includes 8 data points. For an input parameter, reconstructed features can be defined as f1i ,f2i ,⋯,fN i , where N is the total data point, i represents the number of input features, and i = 1, 2, ⋯,8. For each historical time-series, e.g., decomposed IMF components, the data points were rearranged into an RGB image, with 8-bit depth.

How can I label this image and give it the corrosepnding power values?

[Divyansh]

Dear Danya,

I understand that you have an RGB image with 8 stripes in it. Each stripe represents an hour long record of power values from the power plant. You have N such number of images which you want to input in a CNN network and do your further operations. You want to label each image with their corresponding power values.

Assuming my understanding is correct, I still have the following doubts:

• Since each stripe represents power data for 1 hour, is there a single averaged-out power value that you want to label for that particular hour or there are multiple power values within one hour stripe?
• If it is a prediction task, what exactly do you want to predict?

Can you please clarify these doubts?

[Danya]

Dear Divyansh,

Yes, your understanding was correct.

To clarify your doubts:

1- Each stripe represents a single non-averaged value of power (1 stripe represents the actual reading from the sensor at a particular hour)

2- It is a prediction task to predict/forecast the future power value using CNN ( Convolutional Neural Networks).

Hope I clarified your doubts.

Please let me know if you need any further elaboration.

Thank you

[Pratham Shah]

Hi Danya,

I have few questions regarding your problem statement.

1- What kid of labels you want to give to your image?

2- As you said, you want to predict the future power value; which parameter are you going to give in the ML model for prediction?

3- If your model is going to predict the power value based on some numerical data, can't you use regression technique? Means is it necessary to use CNN?

[Danya]

Hi Pratham,

1- I want to label it in such a way that each colour represents a power value. So when the labelled image is used as an input to the CNN, the CNN can read 8 different power value for each image.

2- I will predict the future power by using the labelled images as an input to CNN

3- It is possible to use regression (tried that before) but my task is to test if image-based prediction using CNN is more accurate than the regression thechnique, since CNN works well with images

Answer 1

[Shreeya]

Hello Danya

To my understanding, you want to forecast power outputs represented as RGB values. Instead of labelling the dataset, you can use LSTMs since they are best suited for forecasting. A possible solution is:

• Using a CNN Network to extract the essential image features and output a 1D vector, followed by an LSTM network.
• Image Regeneration using Up Convolution method.
• Segment the image vertically in 8 parts to predict the power values based on the colour gradient.

You can refer to the LSTM, CNN and up convolution documentations linked below respectively.

Long short-term memory (LSTM) layer for recurrent neural network (RNN) - MATLAB - MathWorks India

Train Convolutional Neural Network for Regression - MATLAB & Simulink - MathWorks India

Transposed 2-D convolution layer - MATLAB transposedConv2dLayer - MathWorks India

Hope this helps.