The second year plans to be fully content with Multitasking Design .docx

The second year is fully planned with Multitasking Design

target

On the basis of the first year, in the second year, it is necessary to integrate image classification and power generation prediction (Regression) to establish a multi-task learning (MT) model

Dust Classification using YOLO

Regression with Penalty Function

Through the Penalty Loss Function, the rationality of power generation prediction is improved

Research Methods and Procedures

The second year plan can be broken down into four core parts:

2.1 Multitasking (MT) Modeling

Core Concept: Classification and Regression are trained at the same time.

Technology Choices:

Use a shared backbone (e.g., CNN/Transformer) to work on two tasks at the same time

Input: Solar panel image + related sensing data

Output 1 (Classification): Whether the solar panel is clean or dirty (Classification Task).

Output 2 (Regression): Predicts the amount of electricity generated by the solar panel (Regression Task).

Using the Joint Loss Function

$L_{} =α L_{} +β L_{} + Υ L_{}$

Thereinto:

$L_{}$ : Cross Entropy Loss for image classification

$L_{}$ : Mean Square Error (MSE) of regression of energy production

: Penalty function to ensure that the classification results do not contradict the yield forecast

α, β, γ is a weighted hyperparameter

2.2 YOLO Dust Classification

Objective: Label the dirty areas on the solar panels and calculate the percentage of dirt

Method:

Annotation

Detect the label dirt range through YOLO objects (Ali provides Annotated Dataset).

Train the YOLO model to recognize dirt vs. clean

Dirt ratio calculation

Through the YOLO detection results, the number of dirty pixels/the overall area is calculated

Set a threshold to distinguish between light/moderate/severe contamination

This data affects the Penalty Function

2.3 Penalty Loss Function

Objective: Increase penalties when the predicted energy generation is contradictory to the classification results

Design Logic:

If it is classified as Clean, but the forecast power generation is too low→ penalty

If it is classified as Dirty, but the predicted power generation is too high→ penalty

Mathematical formula: r

The projected power generation is corrected by the proportion of dirt d calculated by YOLO $p_{}$

Set an average standard value $p_{}$ for electricity generation and increase the penalty if the results are contradictory:

λ is the weight that adjusts the degree of impact of the punishment

d is the proportion of dirt

$p_{}$ It is the average power generation in a clean state

2.4 Computer Simulation

Goal: Verify that the Multitasking design is superior to traditional Regression

Comparison :

Baseline (first-year model): Simple regression predicts power generation

Multitasking Model: Integrate Classification + Regression

Add Penalty to the modified model

Evaluation pointer :

MAE (Mean Absolute Error) tests the accuracy of predicting power generation

F1-score (Categorical Outcome Evaluation) tests the accuracy of YOLO

Loss Function Analysis Checks if Penalty improves model performance