Recycling Price Prediction for Retired Electric Vehicle Battery Packs

I investigate how to predict the recycling price of a retired electric vehicle battery pack before dismantling by using surface-level data that can be collected rapidly at a recycling site. In my research, I treat the retired electric vehicle battery pack as an object whose remaining value must be estimated under uncertainty, incomplete information, and market volatility. The central motivation is practical: if a recycling enterprise can estimate the price of a retired electric vehicle battery pack accurately at the point of collection, it can avoid unnecessary transportation, reduce dismantling costs, improve sorting decisions, and protect the economic benefits of large-scale recycling. I focus on surface factors such as capacity, service time, cycle count, appearance wear, maintenance frequency, health indication, and market-related indicators such as metal recovery prices and battery price indices. Because these surface factors are noisy, correlated, and only weakly related to the final recycling price, I construct a bidirectional denoising learning framework for feature extraction and combine it with an optimized support vector regression model.

The growth of electric mobility has made the retirement wave of the electric vehicle battery pack a major resource and environmental issue. A retired electric vehicle battery pack still contains valuable metals and functional materials, but its value depends on chemistry, degradation, market demand, policy conditions, and processing feasibility. If I can predict the recycling price of a retired electric vehicle battery pack from easily observable surface data, I can support rapid quotation, pre-screening, and collection decisions. However, the relationship between surface indicators and recycling price is not simple. Many variables interact nonlinearly, and measurement noise can weaken the signal. Therefore, my work combines denoising representation learning with parameter-optimized regression to improve accuracy and generalization across different types of retired electric vehicle battery packs.

1. Research Context and Practical Motivation

I begin from the observation that recycling decisions for a retired electric vehicle battery pack are often made before full dismantling and laboratory testing. At this stage, the decision maker has access mainly to surface data. Such data include the external condition of the electric vehicle battery pack, recorded usage history, maintenance events, and market indicators. These variables can be collected quickly, but they do not directly reveal internal electrochemical states. Consequently, the prediction problem is characterized by weak supervision, noise, redundancy, and nonlinearity. I address this problem by learning a lower-dimensional representation that preserves the most relevant information for recycling price prediction.

The recycling price of a retired electric vehicle battery pack can be influenced by two broad groups of factors. The first group is related to the battery itself, including capacity, service time, cycle count, appearance wear, repair frequency, and health degree. The second group is related to the market, including nickel, cobalt, manganese, and lithium recovery prices, the number of retired batteries, the price index of new power batteries, and the price index of retired power batteries. I use correlation analysis to reduce the initial candidate set and then apply bidirectional denoising learning to extract robust features from the retained variables.

Dimension Typical Surface Data Why It Matters for a Retired Electric Vehicle Battery Pack
Battery identity Capacity, model scale, initial specification Provides a baseline value and indicates the amount of recoverable material or reusable capacity.
Usage history Service time, cycle count, charging depth Reflects aging and remaining useful life potential.
Physical condition Appearance wear, shell integrity, label completeness Indicates safety risk, dismantling difficulty, and second-life feasibility.
Reliability history Maintenance frequency, fault records Signals reliability and expected reutilization cost.
State indication Health degree, capacity retention proxy Summarizes the internal condition of the electric vehicle battery pack.
Market signals Nickel, cobalt, manganese, lithium recovery prices Determines the material recovery value of the electric vehicle battery pack.
Supply and demand Retired battery quantity, new battery price index, retired battery price index Captures market tightness and expected recycling price movement.

I emphasize that the recycling price of a retired electric vehicle battery pack is not determined by a single variable. For example, a high-capacity electric vehicle battery pack with heavy wear may have a lower price than a smaller electric vehicle battery pack with mild wear and high health degree. Similarly, a rise in lithium or cobalt recovery prices can increase the value of a retired electric vehicle battery pack even when its physical condition is unchanged. This multi-factor dependency motivates my use of feature extraction and nonlinear regression.

2. Problem Characterization of Surface-Data-Based Price Prediction

The problem I study can be stated as follows: given a vector of surface indicators for a retired electric vehicle battery pack, estimate its recycling price. The vector includes both battery-specific and market-specific variables. The prediction target is continuous, and the mapping from inputs to price is nonlinear. The main difficulties are summarized below.

Challenge Description Consequence for Prediction
Weak relevance Surface data only indirectly indicate internal state and future reuse value. Direct regression may underfit the true price mechanism.
Noise Measurement errors, recording errors, and market fluctuations contaminate inputs. Model parameters may be biased and predictions unstable.
Redundancy Service time, cycle count, and wear are often correlated. Multicollinearity can reduce interpretability and generalization.
Nonlinearity Interactions among capacity, health, market prices, and demand are complex. Linear models cannot capture the price response well.
Heterogeneity Different electric vehicle battery pack types have distinct usage patterns and data distributions. A single fixed model may fail across battery types.
Data scarcity Recycling is still developing, and labeled cases may be limited for some types. Deep models with many parameters may overfit.

To reduce the initial factor set, I use Spearman rank correlation. This method is suitable because it captures monotonic relationships and is less sensitive to outliers than ordinary linear correlation. I compute the rank correlation between each candidate surface factor and the recycling price of the retired electric vehicle battery pack. The coefficient is defined as:

$$ r_s = 1 – \frac{6\sum_{i=1}^{n}(x_i-y_i)^2}{n(n^2-1)} $$

where n is the number of samples, x_i is the rank of a candidate factor for sample i, and y_i is the rank of the recycling price for sample i. A value close to 1 or -1 indicates a strong monotonic relationship, while a value close to 0 indicates weak relevance. Based on this analysis, I retain 13 surface factors that cover both the internal condition of the electric vehicle battery pack and external market conditions.

Category Factor Unit or Form Expected Relationship with Recycling Price
Battery self-factor Battery capacity kWh Higher capacity generally increases recoverable value.
Battery self-factor Service time Hours Longer service time generally lowers remaining value.
Battery self-factor Cycle count Times More cycles indicate more aging and lower price.
Battery self-factor Appearance wear Ordinal score Higher wear increases risk and reduces price.
Battery self-factor Maintenance frequency Times More repairs may signal lower reliability.
Battery self-factor Health degree Percentage Higher health increases reuse potential and price.
Market factor Nickel recovery price Currency per kg Higher metal price tends to increase material recovery value.
Market factor Cobalt recovery price Currency per kg Higher cobalt price tends to raise recycling price.
Market factor Manganese recovery price Currency per kg Moderate positive effect on material value.
Market factor Lithium recovery price Currency per kg Strong positive effect on recycling economics.
Market factor Retired battery quantity Count Higher supply may reduce price under fixed demand.
Market factor New power battery price index Percentage Signals substitution and market expectations.
Market factor Retired power battery price index Percentage Directly reflects trend in retired electric vehicle battery pack values.

I also normalize all factors before learning. The min-max normalization I use is:

$$ x’ = \frac{x-\min(x)}{\max(x)-\min(x)} $$

This step is important because the original variables have different scales. For example, service time is recorded in hours, while price indices are percentages. Without normalization, the learning process would be dominated by large-magnitude variables. After normalization, I train a bidirectional denoising autoencoder to learn compact features and suppress noise. Then I use an optimized support vector regression model to predict the recycling price of the retired electric vehicle battery pack.

3. Overall Framework

My framework has three main stages. In the first stage, I screen surface factors for the retired electric vehicle battery pack and retain the 13 variables described above. In the second stage, I use a bidirectional denoising autoencoder, abbreviated as BDAE, to extract features that are robust to noise and redundancy. In the third stage, I use grey wolf optimization, abbreviated as GWO, to tune support vector regression, abbreviated as SVR, and produce the final recycling price prediction for the electric vehicle battery pack.

Stage Input Operation Output
Factor screening Candidate battery and market factors Spearman rank correlation and domain reasoning 13 surface factors
Feature extraction Normalized surface factors Bidirectional denoising learning Low-dimensional robust features
Price prediction Extracted features GWO-optimized SVR Predicted recycling price
Evaluation Predicted and actual prices RMSE, MAPE, cross-validation Accuracy and stability assessment

The design logic is straightforward. Surface data for a retired electric vehicle battery pack are easy to obtain but noisy. A plain regression model may treat noise as signal. A plain autoencoder may reconstruct noise as well as useful structure. Therefore, I introduce denoising learning in both forward and reverse directions. The bidirectional structure lets the model exploit correlations from two representations of the same electric vehicle battery pack data and correct the error of a single-direction extraction.

4. Bidirectional Denoising Feature Extraction

4.1 Autoencoder and Denoising Autoencoder

An autoencoder learns a compact representation by encoding input data into a hidden layer and then decoding it back to the original space. For a retired electric vehicle battery pack, the input is the normalized surface factor vector. The encoder is:

$$ H = s_f(W\bar{X}+b) $$

The decoder is:

$$ Y = s_g(W’H+b’) $$

The reconstruction loss is:

$$ L = \frac{1}{N}\sum_{i=1}^{N}\left\|x^{(i)}-\hat{x}^{(i)}\right\|^2 $$

A denoising autoencoder extends this idea by corrupting the input with noise and requiring the model to reconstruct the clean data. This forces the hidden representation to be stable. For my prediction task, the clean target is not a separately measured internal state; rather, the clean signal is the underlying structure shared by the surface variables. By learning to reconstruct from noisy inputs, the encoder suppresses random fluctuations and preserves factors that are systematically related to the recycling price of the retired electric vehicle battery pack.

4.2 Bidirectional Structure

I construct a bidirectional denoising autoencoder with a forward path and a reverse path. The forward hidden representation and reverse hidden representation are:

$$ \vec{H}_t = \sigma(\vec{x}_{capacity},\vec{x}_{time},\vec{x}_{health},\dots) $$

$$ \overleftarrow{H}_t = \sigma(\overleftarrow{x}_{price},\overleftarrow{x}_{appearance},\overleftarrow{x}_{price\_index},\dots) $$

The two hidden representations are fused during decoding:

$$ Y_t = \sigma(w_1\vec{H}_t + w_2\overleftarrow{H}_t + b) $$

Because the two paths process the same electric vehicle battery pack sample from different orderings or views, the fused representation reduces the risk that one directional extraction will overemphasize accidental correlations. The hidden layer with the lowest reconstruction error is treated as the extracted feature. I use the exponential linear unit as the activation function because it improves gradient flow and provides robustness in noisy learning tasks. The training updates are:

$$ w_{t+1}=w_t-\alpha\frac{\partial L}{\partial w_t} $$

$$ b_{t+1}=b_t-\alpha\frac{\partial L}{\partial b_t} $$

and the mean squared error is:

$$ MSE = \frac{1}{n}\sum_{i=1}^{n}(\hat{y}_i-y_i)^2 $$

4.3 Network Configuration and Training Procedure

I set the input dimension to 13 because 13 surface factors are retained. I use two hidden layers with 10 and 8 units, and I extract a 6-dimensional feature representation. The batch size is 32, and training runs for up to 2000 iterations. The learning rate is set to a small value to stabilize convergence. The training process is summarized below.

Parameter Value Reason
Input nodes 13 Matches retained surface factors for the electric vehicle battery pack.
First hidden layer 10 Captures primary nonlinear interactions.
Second hidden layer 8 Compresses correlated information.
Feature nodes 6 Provides a compact representation for SVR.
Batch size 32 Balances training stability and memory use.
Iterations 2000 Allows reconstruction error to converge.
Activation ELU Reduces gradient vanishing and improves noise robustness.
Loss MSE Directly measures reconstruction quality.
Step Operation
1 Initialize weights, biases, and learning rate.
2 Normalize surface data of the retired electric vehicle battery pack.
3 Add noise and encode the input into forward and reverse hidden representations.
4 Fuse the hidden representations and decode to reconstruct the original input.
5 Compute reconstruction loss between the original and reconstructed data.
6 Update parameters by gradient descent.
7 Repeat until the loss converges.
8 Use the trained encoder to extract features for price prediction.

4.4 Feature Extraction Performance Test

To test the feature extraction capability independently, I use a public housing price dataset as a benchmark because it contains multiple correlated attributes and a continuous target. I use 14 features and compare reconstruction error under different training and testing splits. The purpose is not to predict housing prices but to verify whether the bidirectional denoising structure can learn low-dimensional representations from multi-attribute data. The reconstruction errors remain low across different splits, which indicates stable feature learning.

Training:Testing Split Training Error Testing Error
6:4 0.0067 0.0078
7:3 0.0057 0.0064
8:2 0.0048 0.0055
9:1 0.0081 0.0082

I also compare the bidirectional denoising autoencoder with a deep belief network and a convolutional neural network. Under the same data conditions, the BDAE achieves lower reconstruction error. This supports my decision to use it for the retired electric vehicle battery pack data, where noise and correlated features are both present.

Method Split Reconstruction Error
BDAE 6:4 0.0051
BDAE 7:3 0.0048
BDAE 8:2 0.0045
BDAE 9:1 0.0050
DBN 7:3 0.0180
CNN 7:3 0.0230

5. GWO-Optimized Support Vector Regression

5.1 Support Vector Regression Formulation

After feature extraction, I use support vector regression to map the low-dimensional features to the recycling price of the retired electric vehicle battery pack. SVR is suitable because it can model nonlinear relationships, performs well with limited samples, and has a regularization mechanism that controls overfitting. The regression function is:

$$ y = W\cdot \phi(x)+b $$

The primal optimization problem is:

$$ \min_{W,b}\frac{1}{2}\|W\|^2 + C\sum_{i=1}^{l}(\xi_i+\xi_i^*) $$

Subject to prediction error bounds and slack variables, the dual form becomes:

$$ \max\left[-\frac12\sum_{i=1}^m\sum_{j=1}^m(a_i-a_i^*)(a_j-a_j^*)K(x_i,x_j)-\sum_{i=1}^m a_i^*(y_i+\epsilon)+\sum_{i=1}^m a_i(y_i-\epsilon)\right] $$

The prediction is:

$$ y_i = \sum_{i\in SV}(a_i-a_i^*)K(x_i,x_j)+b $$

I select the radial basis function kernel because it can represent complex nonlinear responses and has fewer numerical difficulties than high-order polynomial kernels. The kernel is:

$$ K(x_i,x_j)=\exp\left(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\right) $$

Kernel Expression Suitability for Retired Electric Vehicle Battery Pack Price
Linear K(x_i,x_j)=x_i^T x_j Too restrictive because surface factors and price are not linearly related.
Polynomial K(x_i,x_j)=(\gamma x_i^T x_j+r)^d Can overfit and becomes costly as degree increases.
Radial basis function K(x_i,x_j)=exp(-||x_i-x_j||^2/(2\sigma^2)) Strong nonlinear mapping and good stability for heterogeneous battery data.

5.2 Grey Wolf Optimization

The performance of SVR depends on the penalty parameter C and the kernel width sigma. Instead of selecting them by experience, I use grey wolf optimization. In my formulation, each wolf represents a candidate pair of C and sigma. The wolves search the parameter space, and the prey is the best recycling price prediction performance for the retired electric vehicle battery pack. The best three candidates are named alpha, beta, and delta, and the rest are omega. The position update equations are:

$$ C=2r_1 $$

$$ a=2-2\frac{t}{T_{max}} $$

$$ A=2a r_2 – a $$

$$ D=|C X_p(t)-X(t)| $$

$$ X(t+1)=X_p(t)-A D $$

where r_1 and r_2 are random numbers in [0,1], t is the current iteration, and T_max is the maximum number of iterations. The final updated position is obtained by averaging the guidance from alpha, beta, and delta. I use ten-fold cross-validation to evaluate each candidate. The cross-validated coefficient of determination is:

$$ R^2_{cv}=1-\frac{\sum_{j=1}^k RSS_j}{\sum_{j=1}^k TSS_j} $$

where:

$$ RSS_j = \sum_{i\in X^j_{test}}(y_i-\hat{y}_i)^2 $$

$$ TSS_j = \sum_{i\in X^j_{test}}(y_i-\bar{y}^j_{test})^2 $$

$$ \bar{y}^j_{test}=\frac{1}{n_j}\sum_{i\in X^j_{test}} y_i $$

A higher cross-validated R-squared indicates that the parameter pair produces a better and more stable price prediction for the retired electric vehicle battery pack. This adaptive search improves generalization across different battery types, which is essential because electric vehicle battery pack models differ in capacity, usage history, and market positioning.

Optimization Component Setting or Meaning
Search variables C and sigma
Lower bound 0.01
Upper bound 20000
Number of optimized variables 2
Maximum iterations 20
Validation Ten-fold cross-validation
Target Maximum cross-validated R-squared

5.3 Prediction Workflow

The workflow I use for the recycling price of a retired electric vehicle battery pack is as follows. First, I collect and clean surface data. Second, I normalize the 13 factors. Third, I train the BDAE and extract six-dimensional features. Fourth, I initialize the wolf population and evaluate each candidate C and sigma using SVR and cross-validation. Fifth, I update the wolf positions until the maximum iteration is reached. Sixth, I train the final SVR with the best parameters. Finally, I output the predicted recycling price.

Step Action Result
1 Read battery and market data Raw records for each electric vehicle battery pack
2 Normalize features Comparable scales
3 Train BDAE Denoised low-dimensional features
4 Initialize GWO Candidate parameter pairs
5 Evaluate fitness Cross-validated performance
6 Update wolves Improved parameter pairs
7 Train final SVR Price prediction model
8 Predict new cases Recycling price for the electric vehicle battery pack

5.4 Evaluation Metrics

I evaluate prediction accuracy using root mean square error and mean absolute percentage error. I also use cross-validated R-squared during parameter optimization. The formulas are:

$$ RMSE = \sqrt{\frac{1}{N}\sum_{i=1}^{N}|y_i-\hat{y}_i|^2} $$

$$ MAPE = \frac{100\%}{n}\sum_{i=1}^{n}\left|\frac{y_i-\hat{y}_i}{y_i}\right| $$

RMSE measures absolute error and is sensitive to large deviations. MAPE measures relative error and is independent of scale. Lower values indicate better prediction of the recycling price for the retired electric vehicle battery pack.

6. Development of a Recycling Price Prediction Support System

I designed a support system to make the method usable in an actual recycling environment. The system follows a browser-server architecture and uses PyQt5 for the user interface. It contains four main modules: login, data center, configuration center, and control panel. The system allows a user to enter surface data for a retired electric vehicle battery pack, run feature extraction and prediction, and display the estimated recycling price. I also include database functions for historical cases, battery self-factors, market factors, and prediction records.

Layer Function Main Content
User layer Interaction Login, data entry, result display, password management
Application layer Computation BDAE feature extraction, GWO-SVR price prediction
Data layer Storage Historical battery data, market data, prediction cases, rule database
Module Purpose Key Functions
Login Access control User authentication, account creation, password reset
Data center Data management Store self-factors, market factors, and historical prediction cases
Configuration center Parameter control Set BDAE nodes, batch size, iterations, learning rate, and GWO bounds
Control panel Prediction execution Input capacity, service time, cycle count, health degree, and obtain price

In the configuration center, I set the BDAE input layer to 13 nodes, the hidden layer to 6 nodes, the batch size to 32, the iteration count to 2000, and the learning rate to 0.0001. The training-to-testing ratio is 8:2. The reconstruction loss is below 0.05, which indicates that the extracted features preserve the main structure of the surface data while reducing noise. For GWO, I set the lower bound to 0.01 and the upper bound to 20000. The algorithm reaches its best accuracy after two iterations in the case study, with an optimization accuracy of 0.813.

System Setting Value
Input layer nodes 13
Hidden layer nodes 6
Batch size 32
Iterations 2000
Learning rate 0.0001
Training set ratio 0.8
Testing set ratio 0.2
GWO lower bound 0.01
GWO upper bound 20000
Optimized variables 2
GWO iterations 20

7. Case Application and Result Analysis

I apply the method to a dataset from a recycling enterprise. The dataset contains 5000 records covering five different models of retired electric vehicle battery packs under the same general specification. The surface variables include capacity, service time, cycle count, maintenance frequency, health degree, nickel price, cobalt price, manganese price, lithium price, retired battery quantity, new battery price index, retired battery price index, appearance wear, and the observed recycling price. I use an 8:2 training-to-testing split after normalization.

Variable Example 1 Example 2 Example 3 Example 4 Example 5 Example 6 Example 7 Example 8
Battery capacity (kWh) 100 85 102 95 70 80 77 65
Service time (h) 34566 45473 42387 42898 38443 36283 37829 42781
Cycle count 7876 7865 6534 8723 8532 9876 5437 6234
Maintenance frequency 3 2 4 1 5 0 0 3
Health degree (%) 81 80 82 75 77 68 85 77
Nickel price 36.25 43.75 45.97 37.36 44.72 44.86 45.00 43.05
Cobalt price 36.94 36.80 36.52 37.36 37.22 37.08 38.09 36.67
Manganese price 18.33 18.88 18.87 17.92 17.90 18.19 17.78 18.89
Lithium price 92.36 91.94 92.91 90.69 91.38 92.98 90.13 90.83
Retired battery quantity 2654 3564 3632 2964 3463 2659 3459 6543
New battery price index (%) 1.1 -2.2 -1.8 0.3 0.9 2.3 -3.6 -0.2
Retired battery price index (%) -2.2 -3.1 -0.4 0.9 3.1 1.3 -1.4 1.2
Appearance wear score 5 1 1 1 5 3 1 2
Recycling price 1354.56 898.86 1006.92 907.91 863.89 937.68 1120.21 1243.91

The prediction results show that the model follows the general trend of the observed recycling price of the retired electric vehicle battery pack. Some peak values are underestimated because surface data cannot fully reveal internal electrochemical states or hidden maintenance history. Nevertheless, most predicted values are close to the actual values, which indicates that the surface data still contain useful information for price estimation. The BDAE-GWO-SVR method performs better than direct SVR and GWO-SVR, and it also outperforms random forest and back-propagation neural network baselines in my comparison.

Model RMSE MAPE Observation
SVR 0.453 0.073 Direct use of surface factors leaves noise and redundancy.
GWO-SVR 0.413 0.067 Parameter optimization improves stability.
BDAE-GWO-SVR 0.371 0.060 Best result because denoising and parameter optimization are combined.
Random forest 1.058 0.084 Weaker for this small and noisy surface-data setting.
Back-propagation neural network 0.812 0.080 Requires more data and is less stable here.

The improvement from BDAE-GWO-SVR comes from two sources. First, the BDAE adjusts the weights of the surface factors and suppresses redundant or noisy variations, which makes the extracted features more informative for the recycling price of the retired electric vehicle battery pack. Second, the GWO searches for the penalty parameter and kernel width of SVR, which allows the regression boundary to adapt to different electric vehicle battery pack models. This combination is important because battery heterogeneity is high: two electric vehicle battery packs with similar capacity can have different service histories and different market values.

Comparison Aspect Direct SVR GWO-SVR BDAE-GWO-SVR
Noise handling Weak Weak Strong through denoising learning
Redundancy handling Weak Weak Strong through compact hidden representation
Parameter adaptation Manual or fixed Automatic Automatic
Generalization across models Limited Improved Best in my case study
Use of market and battery factors Direct Direct Learned feature fusion

8. Discussion

I interpret the results as evidence that surface data can support recycling price prediction for a retired electric vehicle battery pack when the model is designed to handle noise and heterogeneity. A direct regression model assumes that each surface variable contributes independently and consistently. In reality, the variables interact. For instance, cycle count and service time jointly affect health, and health interacts with metal prices to determine the economic value of the electric vehicle battery pack. The BDAE learns a representation that captures these interactions, while the GWO-SVR predictor maps the representation to price with adaptive nonlinearity.

I also note that the prediction target is an economic outcome, not a pure physical measurement. Market sentiment, policy changes, local supply, and contractual conditions can shift the recycling price of a retired electric vehicle battery pack after the data are collected. Therefore, even a strong model should be used as a decision aid rather than an unchallengeable quotation. The support system I designed reflects this view by combining prediction with data management and configuration control. A user can update market indicators and rerun the model, which makes the system more responsive than a static formula.

Strength Explanation Practical Benefit
Bidirectional denoising Uses two hidden paths to correct single-direction error. More robust features for noisy recycling records.
Compact representation Reduces 13 factors to 6 features. Less redundancy and faster downstream training.
Adaptive SVR Optimizes C and sigma by GWO. Better fit across different electric vehicle battery pack models.
Support system Integrates data, parameters, and prediction. Easier deployment at a recycling site.
Limitation Explanation Possible Improvement
Surface data only Internal electrochemical state is not directly measured. Add rapid electrochemical impedance or capacity check when available.
Dynamic market Metal prices and policy signals change quickly. Introduce real-time monitoring and attention-based updating.
Limited interpretability SVR and autoencoders are not fully transparent. Use SHAP or LIME to explain feature contributions.
Data imbalance Some electric vehicle battery pack types may have fewer records. Apply transfer learning or data augmentation.

9. Conclusion and Future Work

I have presented a method for predicting the recycling price of a retired electric vehicle battery pack using surface data. I screened 13 factors from battery self-characteristics and market conditions, used Spearman rank correlation to support factor selection, and then applied a bidirectional denoising autoencoder to extract robust low-dimensional features. I used grey wolf optimization to tune support vector regression so that the model can adapt to different electric vehicle battery pack types. In my case study, the proposed BDAE-GWO-SVR model achieved an RMSE of 0.371 and a MAPE of 0.060, outperforming direct SVR, GWO-SVR, random forest, and back-propagation neural network baselines.

My main conclusion is that bidirectional denoising learning is effective for surface-data-driven recycling price prediction because it reduces noise and redundancy before regression. The optimized SVR then models the nonlinear relationship between the extracted features and the recycling price of the retired electric vehicle battery pack. The support system demonstrates that the method can be integrated into a practical workflow with data entry, parameter configuration, prediction, and result display. This is useful for recycling enterprises that need fast quotations before transportation and dismantling.

For future work, I will extend the model in three directions. First, I will incorporate real-time market monitoring and policy signals so that the predicted price of the retired electric vehicle battery pack can respond quickly to external changes. Second, I will add interpretability tools such as SHAP and LIME so that decision makers can understand which factors drive a specific price prediction. Third, I will improve the human-machine interaction of the support system and collect more data across battery chemistries, formats, and usage histories. These steps will make the prediction of recycling prices for the retired electric vehicle battery pack more accurate, transparent, and practical.

Contribution Method Result
Surface factor screening Spearman rank correlation and domain analysis 13 factors retained for the electric vehicle battery pack.
Noise-robust feature extraction Bidirectional denoising autoencoder Low reconstruction error and compact six-dimensional features.
Adaptive price prediction GWO-optimized SVR RMSE 0.371 and MAPE 0.060 in the case study.
Practical deployment Browser-server support system with PyQt5 interface Rapid input, prediction, and result display for a retired electric vehicle battery pack.

Overall, I believe that predicting the recycling price of a retired electric vehicle battery pack from surface data is feasible when denoising representation learning and parameter-optimized regression are combined. The method does not replace detailed testing, but it provides a fast and useful first estimate. For a recycling enterprise, this estimate can support collection decisions, logistics planning, and value assessment. For the broader circular economy, it can help direct retired electric vehicle battery packs toward appropriate reuse or material recovery pathways.

Scroll to Top