Under the dual-carbon background, China is undergoing a comprehensive green transformation to promote a low-carbon, green, and circular development path. The new energy vehicle industry has emerged as a strategic emerging sector and a key driver of high-quality economic growth. In recent years, China’s new energy vehicle industry has entered a golden period, leading the global market and continuously enhancing its influence and competitiveness. The EV battery pack, as the primary power source for new energy vehicles, has been widely adopted due to its lightweight design, high energy density, long cycle life, and superior safety. With the rapid growth of patent applications and transaction volumes related to the EV battery pack, there is an urgent need to evaluate its value accurately and objectively. However, the complexity of factors affecting the value of EV battery pack patents makes traditional evaluation methods inadequate. Thus, exploring a more effective evaluation approach is essential.
This article focuses on the patent value evaluation of the EV battery pack. I first introduce the relevant concepts of patent evaluation and analyze the difficulties in evaluating EV battery pack patents and the applicability of traditional methods. I then propose a neighborhood rough set-improved BP neural network model to enhance evaluation accuracy. A preliminary indicator system is constructed from four dimensions: textual, technical, legal, and market value. After attribute reduction via neighborhood rough set theory, a final scientific and objective evaluation index system is established. To properly assess the value of EV battery pack patents, I employ a BP neural network model and introduce the gray wolf optimization (GWO) algorithm and AdaBoost algorithm to formulate the GWO-BPNN-AdaBoost model. The performance of the BPNN model, the BPNN-AdaBoost model, and the proposed GWO-BP-AdaBoost model are compared. Empirical results show that the proposed model achieves a prediction accuracy of 91.67%, lower mean absolute percentage error, and a significantly improved R2, verifying its validity and superiority for EV battery pack patent valuation.
1. Introduction and Literature Review
The assessment of patent value has become a critical issue in intellectual property management. My review of the existing literature indicates that research has mainly progressed from traditional appraisal methods toward more advanced computational models. Traditional cost-based, market-based, and income-based methods are widely used but suffer from inherent drawbacks. For example, the cost method assumes a strong correspondence between R&D cost and patent value, which is often unrealistic for complex technologies such as the EV battery pack. The income method relies heavily on predicting future cash flows, which becomes highly uncertain in a rapidly evolving battery technology environment. The market method requires an active market with sufficient comparable transactions, yet the innovation and confidentiality of EV battery pack patents often impede such market development.
Recent scholarship has introduced comprehensive evaluation methods such as the analytic hierarchy process (AHP), fuzzy comprehensive evaluation, entropy weight methods, and more advanced machine learning algorithms. Decision trees, random forests, and neural networks have become increasingly popular for patent valuation due to their ability to capture complex nonlinear relationships. Among these, the BP neural network has gained prominence as one of the most stable and effective models. Yet, conventional BP neural networks have limitations, including slow convergence and sensitivity to initial weights, which can lead to suboptimal local minima. Hence, I integrate a neighborhood rough set for data reduction, the gray wolf optimizer for initial weight optimization, and the AdaBoost algorithm for ensemble boosting.
Focusing on the EV battery pack, the current literature is dominated by patent analysis studies rather than valuation. Several scholars perform technology landscaping of power battery patents, but only a small number of papers address value assessment. Given the increasing importance of EV battery pack patents in corporate financing, licensing, and cross-border transactions, I identify a gap in research and propose a novel machine learning framework that is both objective and accurate.
2. Theoretical Background and Methodology
2.1 Patent Value and the EV Battery Pack
The EV battery pack is the core energy storage component in electric vehicles, with technologies spanning materials, cell design, battery management systems, and integration. These patents exhibit high technical complexity, rapid iteration, and strong coupling with environmental and industrial policies. The value of EV battery pack patents derives not only from direct licensing or transfer income, but also from strategic advantages such as market exclusivity and technological leadership. However, the assessment is complicated by the pace of technology change and the potential substitution of solid-state batteries or sodium-ion cells.
2.2 Limitations of Traditional Evaluation Approaches
I critically evaluate the applicability of three traditional patent appraisal methods to EV battery pack patents:
- Cost Approach: The cost of developing an EV battery pack patent is weakly related to its market potential. R&D costs reflect input but not output value; moreover, many innovations arise from proprietary, uncodified knowledge.
- Income Approach: Estimating the excess income attributable to a specific EV battery pack patent is extremely uncertain. Dynamic changes in battery chemistry, raw material prices, and downstream competition challenge stable cash-flow forecasting.
- Market Approach: The absence of a well-established public market for EV battery pack patents and limited disclosed transaction data make it difficult to obtain reliable market comparables.
Thus, a data-driven, machine-learning-based approach is more suitable for capturing the multidimensional determinants of EV battery pack patent value.
2.3 Neighborhood Rough Set Theory
Rough set theory, originally proposed by Pawlak, enables the discovery of knowledge from imprecise data through attribute reduction. However, standard rough sets require discretization of continuous attributes, which may result in information loss. The neighborhood rough set theory introduced by Hu and colleagues addresses this issue by defining similarity based on the Euclidean distance between continuous attribute values. Let the universe be:
$$U = \left\{ {{x_1},{x_2}, \ldots ,{x_n}} \right\}$$
For any sample x on a nonempty finite set Ω with a neighborhood relation N, the neighborhood of x is defined:
$$\delta (x) = \left\{ {{x_i} \in U\mid \Delta (x,{x_i}) \leqslant \delta } \right\}$$
where Δ is a distance metric and δ is the neighborhood radius. The lower approximation of decision class D with respect to condition attribute subset B is used to compute the dependency degree:
$$\gamma _B (D) = \frac{{|{\underline{N_B} D}|}}{{|U|}}$$
Attribute importance is measured by the change in the dependency function when an attribute is added to the current reduction set. A forward greedy search algorithm generates the final reduced subset while preserving classification ability.
2.4 BP Neural Network
A BP neural network is a multilayer feedforward network trained by error backpropagation. Given an input vector x, the hidden-layer output is:
$$h_j = f\left( {\sum_{i = 1}^n {w_{ij} x_i + b_j} } \right)$$
and the output-layer result is:
$$y_k = g\left( {\sum_{j = 1}^m {w_{jk} h_j + b_k} } \right)$$
The loss function minimizes the squared error between the actual output and target value. The iterative process continues until the loss is below a predefined threshold or the maximum number of epochs is reached. In this study, the BP model has three hidden layers with 600, 800, and 1200 neurons, respectively.
2.5 Grey Wolf Optimizer
The grey wolf optimizer mimics the leadership hierarchy and hunting mechanism of grey wolves. The pack is divided into alpha (α), beta (β), delta (δ), and omega (ω) wolves. The main equations for encircling behavior are:
$$D = \left| {C \cdot X_P (t) – X(t)} \right|$$
$$X(t + 1) = X_P (t) – A \cdot D$$
where A and C are coefficient vectors, and XP is the position of the prey. The coefficients are calculated as:
$$A = 2a \cdot r_1 – a,\quad C = 2 \cdot r_2$$
Here, r1 and r2 are random vectors in [0,1], and a linearly decreases from 2 to 0. The GWO algorithm efficiently searches for the optimal initial weights and thresholds of the BP network, preventing premature convergence.
2.6 AdaBoost Algorithm
AdaBoost is an ensemble method that combines multiple weak learners sequentially. Let T be the number of weak learners. Initially, sample weights are set uniformly as:
$$w_1 (i) = \frac{1}{n},\quad \forall i = 1, \ldots ,n$$
The training procedure generates a weak model Gt at iteration t based on the weighted samples. The weighted error rate is computed by:
$$e_t = \sum_{i = 1}^n w_t (i) \cdot I\left( {{G_t} (x_i) \ne y_i} \right)$$
The model weight is:
$$\alpha _t = \frac{1}{2}\ln \left( {\frac{{1 – e_t}}{{e_t}}} \right)$$
and the sample weights are updated. The final strong predictor is a weighted combination:
$$G(x) = {\mathop{\mathrm{sign}}\nolimits} \left( {\sum_{t = 1}^T {\alpha _t \cdot {G_t} (x)} } \right)$$
3. Construction of Evaluation Index System for EV Battery Pack Patents
3.1 Initial Index System
Drawing on prior research and the specific traits of the EV battery pack, I constructed an initial index system with four primary dimensions: textual value, technical value, legal value, and market value. Each dimension contains a set of measurable indicators. Table 1 displays the complete set of 18 preliminary indicators.
Table 1. Initial patent value evaluation index system for EV battery pack patents
| Primary dimension | Secondary indicator (code) | Meaning | Measurement benchmark |
|---|---|---|---|
| Textual value | Patent text quality (C1) | Theoretical basis and drafting quality | Number of pages in the specification |
| Textual value | Number of drawings (C2) | Patent complexity and clarity | Count of drawing sheets |
| Textual value | Number of inventors (C3) | Technical level and innovation | Count of listed inventors |
| Textual value | Number of claims (C4) | Technical feature quality and scope | Count of claims |
| Technical value | Citations made (C5) | Level of knowledge integration | Count of backward citations |
| Technical value | Citations received (C6) | Technological influence | Count of forward citations |
| Technical value | Patent breadth (C7) | Technical extent and complexity | Number of IPC symbols |
| Technical value | Technical advancement (C8) | Superiority over existing technology | IncoPat composite score |
| Technical value | Technical stability (C9) | Reliability and predictability | IncoPat stability score |
| Legal value | Patent legal status (C10) | Level of protection and validity | Invalid = 1, valid = 2 |
| Legal value | Remaining legal life (C11) | Profit period still available | Remaining years of validity |
| Legal value | Whether pledged/insured (C12) | Independence and encumbrance | Yes = 2, no = 1 |
| Legal value | Patent protection scope (C13) | Market exclusivity and coverage | IncoPat scope score |
| Market value | Applicant type (C14) | Commercialization capability | Individual = 1, enterprise = 2, university = 3 |
| Market value | Patent family size (C15) | Economic potential and global presence | Count of family members |
| Market value | Cooperative Patent Classification (C16) | Market coverage and cross-sector reach | Number of CPC symbols |
| Market value | Number of fee payments (C17) | Willingness to maintain and market prospects | Count of maintenance fee payments |
| Market value | Number of transfers (C18) | Market demand and applicability | Count of assignment/transfer events |
3.2 Data Collection and Preprocessing
I searched for “lithium-ion power battery” patents on reputable intangible-asset trading platforms including Huijia and Yuzhua. From more than 230 records, I obtained the associated patent numbers and retrieved detailed metadata from the IncoPat global patent database. After excluding patents with missing or invalid indicator data, I retained 200 valid patent records for empirical analysis. Among them, 182 are invention patents and 18 are utility models. Because the actual transaction prices of patents are rarely disclosed and listing prices fluctuate widely, I adopted the “IncoPat composite value” as a proxy for the true patent value. This composite value is generated automatically via multi-dimensional scoring and is free of human subjective bias.
Owing to mixed quantitative and qualitative indicators, I applied a standardization process. Qualitative variables, such as whether the patent is pledged, were mapped to numeric codes. For quantitative variables that exhibit skewed distributions, I converted them into ordinal categories from 1 to 5 based on frequency analysis. This data transformation aligns with the requirements of the neighborhood rough set approach and stabilizes neural network training.
Table 2. Standardization criteria for the indicators
| Indicator | Level 1 | Level 2 | Level 3 | Level 4 | Level 5 |
|---|---|---|---|---|---|
| C1: Patent text quality (pages) | [1,7) | [7,10) | [10,13) | [13,16) | >=16 |
| C2: Number of drawings | [0,2) | [2,4) | [4,6) | [6,8) | >=8 |
| C3: Number of inventors | [1,3) | [3,5) | [5,7) | [7,9) | >=9 |
| C4: Claims count | [1,5) | [5,9) | [9,13) | [13,17) | >=17 |
| C5: Backward citations | [0,2) | [2,4) | [4,6) | [6,8) | >=8 |
| C6: Forward citations | [0,2) | [2,4) | [4,6) | [6,8) | >=8 |
| C7: Patent breadth (IPC count) | [1,3) | [3,5) | [5,7) | [7,9) | >=9 |
| C8: Technical advancement | [1,3) | [3,5) | [5,7) | [7,9) | >=9 |
| C9: Technical stability | [1,3) | [3,5) | [5,7) | [7,9) | >=9 |
| C10: Legal status | Invalid/failure | Valid | – | – | – |
| C11: Remaining life (years) | [0,4) | [4,8) | [8,12) | [12,16) | >=16 |
| C12: Pledge status | Yes | No | – | – | – |
| C13: Protection scope | [1,3) | [3,5) | [5,7) | [7,9) | >=9 |
| C14: Applicant type | Individual | Enterprise | University | – | – |
| C15: Family size | [1,2) | [2,3) | [3,4) | >=4 | – |
| C16: CPC count | [0,2) | [2,4) | [4,6) | [6,8) | >=8 |
| C17: Fee payment count | [1,5) | [5,9) | [9,13) | [13,17) | >=17 |
| C18: Transfer count | [0,1) | [1,2) | [2,3) | [3,4) | >=4 |
After categorical discretization, I also normalized all numeric inputs using the min-max scaling method:
$$x_{\text{norm}} = \frac{{x – x_{\min} }}{{x_{\max} – x_{\min} }}$$
This procedure maps each attribute to the range [0, 1], thereby preventing features with larger magnitudes from dominating the neural network training process and accelerating convergence.
3.3 Index Reduction Using Neighborhood Rough Set
In this framework, the 18 indicators (C1–C18) serve as condition attributes, while the IncoPat composite value is the decision attribute. I implemented the neighborhood rough set algorithm in MATLAB. To select an adequate neighborhood radius, I tuned the parameter λ that controls the ratio between the standard deviation of each attribute and the radius:
$$\delta = \frac{{std(A)}}{\lambda }$$
I set the significance cutoff to 0.001 and 0.01 and varied λ between 0.5 and 1.5 with a step size of 0.1. Table 3 shows the resulting attribute subsets for different parameter combinations.
Table 3. Reduced attribute subsets under different parameter settings
| Sig | λ = 0.5 | λ = 1.0 | λ = 1.3 |
|---|---|---|---|
| 0.001 | 1,2,3,4,7,8,10,11,12,13,14,15,16,17 | 1,2,3,4,5,6,7,8,12,13,14,16,17 | 1,4,7,8 |
| 0.01 | 3,4,8,11,12,15,16,17 | 1,3,4,5,6,7,8,13,14,16,17 | 1,4,7,8 |
By counting the frequency of each indicator over all parameter combinations, I retained indicators that appeared more than ten times. Consequently, the final reduced system contains 11 indicators: C1, C2, C3, C4, C5, C7, C8, C11, C12, C16, and C17. The indicators C6, C9, C10, C13, C14, C15, and C18 were removed as redundant for the evaluation of EV battery pack patents in this dataset. The final system is shown in Table 4.
Table 4. Final EV battery pack patent value evaluation index system after neighborhood rough set reduction
| Primary dimension | Secondary indicators | Evaluation codes |
|---|---|---|
| Textual value | Patent text quality, number of drawings, inventor count, claims count | C1, C2, C3, C4 |
| Technical value | Backward citations, patent breadth, technical advancement | C5, C7, C8 |
| Legal value | Remaining legal life, pledge status | C11, C12 |
| Market value | CPC classification count, fee payment count | C16, C17 |
4. Construction and Training of the Improved BP Neural Network Model
4.1 Input and Dataset Splitting
From the 200 valid patent samples, I randomly chose 140 records (70%) as the training set and retained 60 patents (30%) as the test set to evaluate generalization capacity. The input dimension equals the 11 selected indicators, and the output is the single continuous value of the EV battery pack patent value.
4.2 Structural Design of the GWO-BP-AdaBoost Model
The network architecture is a multilayer perceptron with an input layer of 11 nodes, three hidden layers with 600, 800, and 1200 nodes, and an output layer of one node. The transfer function used in hidden nodes is the sigmoid function:
$$f(x) = \frac{1}{{1 + e^{ – x} }}$$
I set the training function to the Levenberg–Marquardt algorithm, which is robust for moderate-scale regression tasks. The maximum number of epochs is 1000; the mean squared error goal is set to 1e-5; the learning rate is 0.0005. To determine the best initial weights and biases, the GWO algorithm is employed before the training stage. The population size is 30 and the maximum number of iterations is optimized to 46, as indicated by the convergence analysis in Figure 1. During the training phase, the fitness value decays steadily until iteration 46, beyond which no significant improvement occurs. Therefore, the iteration count is set to 46.
4.3 Training Results
After training, I computed the regression coefficient of the training dataset. The R-value for the training subset is 0.96531, which demonstrates a very strong linear relationship between the model outputs and the observed IncoPat composite values. The high R-value indicates that the model has learned the underlying pattern from the EV battery pack patent indicators and is suitable for prediction.
I also applied the AdaBoost method to boost the performance of multiple weak BP networks. Adaboost training proceeds in ten sequential rounds, where each round retrains the BP network on updated sample weights. The final ensemble output is weighted linearly, thereby reducing the expected error and improving robustness.
4.4 Prediction Performance
To evaluate the model, I fed the 60 test set samples into the trained GWO-BP-AdaBoost predictor. The predicted value obtained from the network is denormalized and compared with the true composite value. For intuitive interpretation, the prediction is approximated to the nearest integer and contrasted with the actual integer-level patent score.
My results reveal that the forecast is generally consistent with actual values. Of the 60 test patents, 91.67% have relative errors lower than 10%. Thus, the proposed hybrid model is reliable and applicable. It is important to note that small deviations may arise because the value of an EV battery pack patent depends on non-systematic factors, including time-specific negotiation contexts and market liquidity, which cannot be fully captured by static patent bibliometrics.
5. Comparative Analysis of Models
To verify the contribution of each optimization module, I compared the proposed GWO-BP-AdaBoost model with two baselines: the basic BP neural network (BPNN) and the BPNN-AdaBoost ensemble model. All models use identical inputs and datasets. The error metrics chosen for comparison are relative error, mean absolute percentage error (MAPE), and R-squared (R2). The calculation formulas are:
$$MAPE = \frac{1}{n}\sum_{i = 1}^n {\left| {\frac{{y_i – \hat y_i}}{{y_i}}} \right| \times 100\% }$$
$$R^2 = 1 – \frac{{\sum_{i = 1}^n {(y_i – \hat y_i)^2}}}{{\sum_{i = 1}^n {(y_i – \bar y)^2}}}$$
Table 5. Performance metrics for the three neural network models applied to EV battery pack patent value prediction
| Model | Accuracy (relative error < 10%) | MAPE (%) | R² |
|---|---|---|---|
| BPNN | 63.00% | 15.43% | 0.714 |
| BPNN-AdaBoost | 90.00% | 8.71% | 0.832 |
| GWO-BP-AdaBoost | 91.67% | 6.72% | 0.907 |
The baseline BPNN model tends to suffer from local minima and instability: several predicted values deviate markedly from the true values, with some relative errors reaching 0.88, 1.38, and 2.59. When the AdaBoost algorithm is added, the relative error significantly decreases, as most test samples now fall below the 10% threshold; however, one residual error still reaches 0.96. The full hybrid GWO-BP-AdaBoost model yields the smallest relative errors in all test samples, with no error exceeding 0.3. The accuracy rate improves from 63.00% to 91.67%, the MAPE drops from 15.43% to 6.72%, and R² rises from 0.714 to 0.907. These results validate the addition of the grey wolf optimizer to choose optimal network weights and the AdaBoost ensemble to improve generalization on EV battery pack patent data.
6. Case Study and Application
6.1 Background of the Case Company
The case study concerns a prominent Chinese EV battery manufacturer that specializes in the research, development, production, and sale of power and energy-storage lithium-ion batteries. The company has established a dense patent portfolio across the whole EV battery pack value chain, including electrode materials, cell structures, battery management systems, and manufacturing processes. Its innovative capacity is demonstrated by thousands of granted patents and a strong pipeline of pending applications. As an essential supplier to leading automakers, the company holds a top tier position in the Chinese power battery market.
6.2 Selection of Evaluation Targets
I randomly selected five patents related to EV battery pack technologies owned by the case company. These patents cover a soft-pack lithium-ion battery module, rolling equipment for battery current collector foils, a solid-state electrolyte membrane for lithium-sulfur batteries, a lithium-ion battery anode prelithiation apparatus, and a negative electrode plate design. The patent characteristics are presented in Table 6.
Table 6. Summary of case patents and their specifications
| Patent code | Patent name | Text pages | Drawings | Inventors | Claims | Backward citations | IPC count | Technical advancement | Remaining life (years) | CPC count | Fee payments |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Patent A | Soft-pack lithium-ion battery module and connection method | 10 | 5 | 5 | 7 | 1 | 7 | 10 | 13.55 | 3 | 6 |
| Patent B | Rolling and leveling apparatus for lithium-ion battery current collector foil | 9 | 3 | 3 | 8 | 5 | 3 | 5 | 13.27 | 4 | 5 |
| Patent C | Lithium-sulfur battery solid electrolyte membrane, its preparation method, and lithium-sulfur battery | 9 | 2 | 3 | 14 | 0 | 4 | 9 | 15.53 | 9 | 4 |
| Patent D | Lithium-ion battery negative electrode prelithiation device and method | 10 | 4 | 4 | 10 | 4 | 1 | 10 | 12.90 | 2 | 7 |
| Patent E | Negative electrode plate and lithium-ion battery including the same | 15 | 1 | 5 | 10 | 3 | 2 | 8 | 18.20 | 1 | 2 |
Since the network requires normalized input, each indicator was discretized according to Table 2, producing the coded data shown in Table 7.
Table 7. Discretized input data for the case patents
| Indicator | Patent A | Patent B | Patent C | Patent D | Patent E |
|---|---|---|---|---|---|
| C11 | 3 | 2 | 2 | 3 | 4 |
| C21 | 3 | 2 | 2 | 3 | 1 |
| C31 | 3 | 2 | 2 | 2 | 3 |
| C41 | 2 | 2 | 4 | 5 | 5 |
| C51 | 1 | 3 | 1 | 3 | 2 |
| C71 | 4 | 2 | 2 | 1 | 1 |
| C81 | 5 | 3 | 5 | 5 | 4 |
| C111 | 4 | 4 | 4 | 4 | 5 |
| C121 | 2 | 2 | 2 | 2 | 2 |
| C161 | 2 | 3 | 5 | 2 | 1 |
| C171 | 2 | 2 | 1 | 2 | 1 |
I set the evaluation benchmark date to 31 October 2024 and treated the market value type under the assumption of an open market. The model was applied with the same hyperparameter settings as described earlier. Table 8 displays the actual true value, the predicted value output, and the corresponding error for each patent.
Table 8. Evaluation results for the five case patents
| Patent | Actual value | Predicted value | Relative error |
|---|---|---|---|
| Patent A | 10 | 9.273 | 0.073 |
| Patent B | 8 | 8.310 | 0.039 |
| Patent C | 9 | 9.634 | 0.070 |
| Patent D | 9 | 9.108 | 0.012 |
| Patent E | 9 | 8.737 | 0.029 |
The mean relative error across all five target patents is only 4.46%, indicating reliable performance. In particular, Patent A and Patent C show deviations close to 7%, while Patent D exhibits the smallest error of 1.2%. All relative errors remain within the acceptable range of 10%. These findings confirm that the GWO-BP-AdaBoost model can effectively estimate the value of EV battery pack patents and provide a trustworthy basis for future patent transactions, licensing agreements, and asset management decisions.
7. Conclusions and Future Directions
The global shift toward sustainable mobility and cleaner energy has elevated the role of the EV battery pack as a core technological asset. This article proposes a comprehensive framework for the patent valuation of the EV battery pack by combining domain-driven evaluation indices, neighborhood rough set dimensionality reduction, and optimized machine learning techniques. I summarize the main contributions as follows.
First, I establish an initial evaluation index system for EV battery pack patents that incorporates not only standard bibliometric indicators, such as the number of pages, drawings, inventors, claims, citations, and patent family size, but also technical and legal composite indicators, including technical advancement and pledge status. This system captures the multidimensional nature of the EV battery pack patents more accurately than conventional patent scoring systems.
Second, I apply neighborhood rough set theory to reduce the initial index set from 18 to a parsimonious subset of 11 indicators. The reduction process removes redundant and collinear information without harming the discrimination capacity. This leads to a more parsimonious, transparent, and objective evaluation system for EV battery pack patents. The selected indicators cover textual, technical, legal, and market aspects and lay a solid foundation for robust predictive modeling.
Third, I construct an improved BP neural network whose initial weights and thresholds are selected by the grey wolf optimizer and whose performance is further boosted by the AdaBoost ensemble algorithm. This hybrid architecture addresses two common weaknesses of the standard backpropagation network: slow convergence and susceptibility to local optima. The grey wolf optimizer facilitates a global search for suitable network initialization, while AdaBoost sequentially adjusts sample weights so that more difficult samples are recursively emphasized. The empirical experiments on 200 EV battery pack patents show that the proposed GWO-BP-AdaBoost model achieves the highest prediction accuracy (91.67%), the lowest MAPE, and the highest R² among the compared models, consistently outperforming both the plain BPNN and the BPNN-AdaBoost baseline.
The case study on a leading battery manufacturer demonstrates that the model is capable of producing quick and defensible estimates of patent value for EV battery pack technologies. The average relative error for the five selected patents is below 5%, demonstrating the model’s practical value in real-world settings.
Although this study yields satisfying results, a few limitations must be acknowledged. First, the benchmark used as the patent value is the IncoPat composite value instead of an actual arm’s-length transaction price. Although this composite value is objective, it incorporates the database’s proprietary weighting scheme. Once a well-functioning public patent market is formed and transaction prices are transparently disclosed, these actual prices can replace the proxy variable for even more accurate results.
Second, my data set is limited to 200 EV battery pack patents. While this sample is sufficient for model training, a larger data set covering more recent technological developments, international patents, and patent lifespans may improve generalization. Third, the model architecture was tuned via empirical experiments but not exhaustively optimized through a broader search over every possible hyperparameter configuration. Future studies could combine Bayesian optimization with the GWO-BP-AdaBoost structure to further improve generalizability.
Fourth, some important qualitative factors, including market demand dynamics, policy orientation, technology maturity, and competitive rivalry, are not fully representable by the available bibliometric data. Future work may develop systematic questionnaires or expert scoring mechanisms that convert these qualitative aspects into quantitative inputs.
Fifth, the proposed model is tailored to EV battery pack patents. I strongly recommend testing it on other complementary technology domains, such as fuel cells, charging infrastructure, or energy storage systems, to determine its cross-domain transferability and robustness. As technological development in EV battery packs evolves rapidly, regular model retraining with newly published patents and market data will keep the assessment framework current and reliable.
In conclusion, the integrated neighborhood rough set and GWO-optimized, AdaBoost-boosted BP neural network model offers a powerful new tool for the objective and effective evaluation of EV battery pack patents. Through better screening of value indicators and improved model accuracy, this research contributes to more efficient management and commercialization of EV battery pack intellectual property, ultimately facilitating the transition to a low-carbon and innovation-driven transportation ecosystem.