Evaluating Lithium-Ion Power Battery Patent Value with Rough Sets and Optimized BP Neural Networks

In my research, I examine how the value of lithium-ion power battery patents can be assessed more accurately in the context of the global transition toward low-carbon mobility. The electric vehicle battery pack is one of the most important technological and economic components of a modern electric vehicle, and the patent portfolio surrounding that electric vehicle battery pack strongly influences firm competitiveness, technology transfer, licensing revenue, and financing capacity. Because the electric vehicle battery pack combines materials science, electrochemistry, mechanical design, thermal management, and electronic control, its patent value is multidimensional and difficult to quantify with a single traditional appraisal method. I therefore develop an integrated framework that uses neighborhood rough sets for indicator reduction and a hybrid BP neural network optimized by the grey wolf optimizer and AdaBoost for value prediction. The framework is intended to support more objective, data-driven, and scalable evaluation of lithium-ion power battery patents, especially when transaction data are scarce and when the technical boundary of an electric vehicle battery pack is rapidly changing.

1. Research Context and Motivation

The transition toward green transportation has made the electric vehicle battery pack a central object of innovation. A high-performance electric vehicle battery pack determines driving range, charging speed, safety, cycle life, and total vehicle cost. For this reason, patents related to the electric vehicle battery pack are not merely legal instruments; they are strategic assets that can shape supply chains, licensing negotiations, and market entry. When I evaluate a lithium-ion power battery patent, I am not only estimating a legal right. I am also assessing the likelihood that the protected technology will improve the electric vehicle battery pack, reduce costs, increase energy density, or strengthen the competitive position of its owner.

Traditional patent valuation methods include the cost approach, market approach, income approach, and real options approach. Each has a role, but each also has limitations for lithium-ion power battery patents. The cost approach struggles because the relationship between research expenditure and patent value is weak. The market approach struggles because comparable transactions for a specific electric vehicle battery pack patent are rare and often undisclosed. The income approach struggles because future cash flows depend on technical substitution, policy changes, and the pace of innovation in the electric vehicle battery pack. Real options methods can capture flexibility, but they often require assumptions that are difficult to validate in a fast-moving battery market. Consequently, I seek a hybrid method that combines data reduction with machine learning so that the value of a lithium-ion power battery patent can be estimated from observable patent characteristics.

My motivation is both theoretical and practical. Theoretically, I want to improve the selection of patent value indicators and reduce subjective weighting. Practically, I want to provide a model that can screen large numbers of lithium-ion power battery patents and identify those with higher potential value. Such a model can help firms decide whether to maintain, license, transfer, pledge, or litigate a patent related to the electric vehicle battery pack. It can also help research institutions and technology transfer offices prioritize patents that are more likely to be commercialized in the electric vehicle battery pack supply chain.

2. Key Concepts and Theoretical Foundations

2.1 Patent Rights and Patent Value

A patent is an exclusive right granted for an invention that meets novelty, inventiveness, and industrial applicability requirements. Patent value can be understood as the expected economic benefit that the patent can generate, but it also includes technical value and legal value. In my framework, patent value is therefore a composite outcome of text quality, technical strength, legal protection, and market potential. For lithium-ion power battery patents, this composite nature is especially visible because the electric vehicle battery pack is a system-level product. A patent may protect a material, a cell structure, a module design, a thermal management strategy, a battery management algorithm, or a manufacturing process. Each type of patent contributes differently to the performance of the electric vehicle battery pack.

2.2 Characteristics of Lithium-Ion Power Battery Patents

I identify three characteristics that shape the evaluation of lithium-ion power battery patents. First, the technology is highly complex and interdisciplinary. A single patent may combine chemistry, materials engineering, electrical engineering, and mechanical design. Second, the technology iterates rapidly. Improvements in cathode materials, anode materials, separators, electrolytes, and cell-to-pack architectures continuously alter the value of existing patents. Third, the technology is strongly affected by policy and market conditions. Subsidies, emissions regulations, battery recycling rules, and consumer demand for longer-range electric vehicles all influence the commercial value of an electric vehicle battery pack patent.

2.3 Evaluation Difficulties

The first difficulty is that many value drivers are not directly observable. For example, the technical stability of a patent may depend on experimental data that are not fully disclosed. The second difficulty is that transaction data are limited. Because patent transactions involving the electric vehicle battery pack often occur inside large firms or through cross-licensing, the true price is rarely public. The third difficulty is that no unified indicator system has been established for lithium-ion power battery patents. Different analysts may select different indicators and assign different weights, which reduces comparability. I address these difficulties by using neighborhood rough sets for objective attribute reduction and by using a hybrid neural network for nonlinear prediction.

3. Neighborhood Rough Set Theory for Indicator Reduction

Rough set theory is a mathematical tool for handling uncertainty and inconsistency. Classical rough sets require discrete data, but patent indicators are often continuous. Neighborhood rough sets overcome this limitation by using a neighborhood relation instead of an equivalence relation. This makes the method suitable for continuous patent indicators such as citation counts, patent family size, and remaining legal life. I use neighborhood rough sets to remove redundant indicators while preserving the classification ability of the original indicator set.

Let the universe of patent samples be:

$$U=\{x_1,x_2,\ldots,x_n\}$$

For a sample \(x_i\), its neighborhood is defined as:

$$\delta(x_i)=\{x_j \mid x_j \in U, \Delta(x_i,x_j)\le \delta\}$$

where \(\Delta\) is a distance metric and \(\delta\) is the neighborhood radius. For a decision attribute \(D\), the lower approximation of a decision class \(X\) with respect to a condition attribute subset \(R\) is:

$$\underline{N_R}X = \{x_i \mid \delta_R(x_i) \subseteq X\}$$

The dependency degree of \(D\) on \(R\) is:

$$\gamma_R(D)=\frac{|\operatorname{POS}_R(D)|}{|U|}$$

The significance of adding an attribute \(a\) to \(R\) is:

$$SIG(a,R,D)=\gamma_{R\cup\{a\}}(D)-\gamma_R(D)$$

I use a forward greedy search to select attributes with high significance. The process stops when additional attributes no longer improve the dependency degree beyond a threshold. This produces a reduced indicator set that is more compact and less collinear than the original set.

4. BP Neural Network and Its Limitations

A BP neural network is a multilayer feedforward network trained by backpropagation. It can approximate complex nonlinear relationships between patent indicators and patent value. In my model, the input layer receives the reduced patent indicators, the hidden layers extract nonlinear interactions, and the output layer produces the estimated value of the lithium-ion power battery patent.

For a neuron \(j\), the weighted sum is:

$$S_j = \sum_{i=1}^n w_{ij}x_i + b_j$$

The output is:

$$y_j = f(S_j)$$

where \(f\) is an activation function. The mean squared error is:

$$E = \frac{1}{2}\sum_{k=1}^m (y_k – \hat{y}_k)^2$$

Weights are updated by gradient descent:

$$w_{ij}^{(t+1)} = w_{ij}^{(t)} – \eta \frac{\partial E}{\partial w_{ij}^{(t)}}$$

Although BP networks are powerful, they have three weaknesses. First, training can be slow when the search space is large. Second, the network can become trapped in local minima. Third, the initial weights and biases strongly affect performance. For the electric vehicle battery pack patent domain, these weaknesses matter because the relationship between indicators and value is noisy and nonlinear. I therefore introduce the grey wolf optimizer to improve the initial parameters and AdaBoost to improve the ensemble prediction.

5. Grey Wolf Optimizer

The grey wolf optimizer simulates the social hierarchy and hunting behavior of grey wolves. The best solution is called \(\alpha\), the second-best is \(\beta\), the third-best is \(\delta\), and the remaining candidates are \(\omega\). The algorithm updates positions around the three best wolves. The encircling behavior is modeled as:

$$D = |C \cdot X_p(t) – X(t)|$$

$$X(t+1) = X_p(t) – A \cdot D$$

where \(A\) and \(C\) are coefficient vectors. The coefficient \(A\) is:

$$A = 2a \cdot r_1 – a$$

The parameter \(a\) decreases linearly from 2 to 0:

$$a = 2 – 2\left(\frac{t}{T_{\max}}\right)$$

The hunting behavior is represented by the positions of \(\alpha\), \(\beta\), and \(\delta\):

$$D_\alpha = |C_1 X_\alpha – X|,\quad D_\beta = |C_2 X_\beta – X|,\quad D_\delta = |C_3 X_\delta – X|$$

$$X_1 = X_\alpha – A_1 D_\alpha,\quad X_2 = X_\beta – A_2 D_\beta,\quad X_3 = X_\delta – A_3 D_\delta$$

$$X(t+1) = \frac{X_1 + X_2 + X_3}{3}$$

I use the grey wolf optimizer to optimize the initial weights and biases of the BP neural network. This reduces the risk of poor initialization and helps the network converge to a better region of the parameter space.

6. AdaBoost Algorithm

AdaBoost is an ensemble learning algorithm that combines weak learners into a strong learner. It assigns weights to training samples and increases the weights of samples that are misclassified. In my model, each weak learner is a BP neural network optimized by the grey wolf optimizer. The error of the \(t\)-th weak learner is:

$$\epsilon_t = \sum_{i=1}^N w_{t,i} I(y_i \neq h_t(x_i))$$

The weight of the weak learner is:

$$\alpha_t = \frac{1}{2}\ln\left(\frac{1-\epsilon_t}{\epsilon_t}\right)$$

The sample weights are updated as:

$$w_{t+1,i} = \frac{w_{t,i}\exp(-\alpha_t y_i h_t(x_i))}{Z_t}$$

where \(Z_t\) is a normalization factor. The final strong learner is:

$$H(x) = \operatorname{sign}\left(\sum_{t=1}^T \alpha_t h_t(x)\right)$$

By combining multiple optimized BP networks, AdaBoost improves the stability and predictive accuracy of the patent value assessment. This is particularly useful for lithium-ion power battery patents because the value signal is weak and heterogeneous across different technical routes within the electric vehicle battery pack.

7. Proposed Integrated Framework

My proposed framework has four stages. In the first stage, I construct an initial patent value indicator system from four dimensions: text value, technical value, legal value, and market value. In the second stage, I collect patent data from public patent databases and transaction platforms, then preprocess the data through discretization and normalization. In the third stage, I apply neighborhood rough sets to reduce the indicator set. In the fourth stage, I train a GWO-BP-AdaBoost model to predict patent value. The final output is an estimated value score for each lithium-ion power battery patent, which can be used to rank patents, support transaction pricing, or guide portfolio management for the electric vehicle battery pack.

8. Initial Indicator System

The initial indicator system contains 18 secondary indicators under four primary dimensions. I summarize them in Table 1. The indicators are selected because they are observable, comparable, and relevant to the value of an electric vehicle battery pack patent.

Primary Dimension Secondary Indicator Meaning Measurement
Text Value Patent text quality (C1) Theoretical support and writing quality Number of pages
Text Value Number of drawings (C2) Complexity and readability of the patent Number of figures
Text Value Number of inventors (C3) Team size and interdisciplinary nature Number of inventors
Text Value Number of claims (C4) Scope and legal strength of claims Number of claims
Technical Value Forward citations (C5) Knowledge accumulation and innovation breadth Citation count
Technical Value Backward citations (C6) Influence on later technologies Cited count
Technical Value Patent width (C7) Technical breadth Number of IPC classes
Technical Value Technical advancement (C8) Degree of technical improvement Database score
Technical Value Technical stability (C9) Reliability and predictability Database score
Legal Value Legal status (C10) Validity of the patent right Invalid = 1, valid = 2
Legal Value Remaining life (C11) Remaining period of legal protection Years
Legal Value Pledge status (C12) Independence and transferability Yes = 2, no = 1
Legal Value Protection scope (C13) Market exclusivity Database score
Market Value Applicant type (C14) Commercialization capability Individual = 1, firm = 2, university = 3
Market Value Patent family size (C15) International market coverage Number of family members
Market Value CPC classes (C16) Market coverage and cross-domain relevance Number of CPC classes
Market Value Fee payment count (C17) Maintenance willingness and market potential Number of payments
Market Value Transfer count (C18) Commercialization activity Number of transfers

9. Data Collection and Preprocessing

I collected 230 lithium-ion power battery patent records from public patent databases and transaction platforms. After removing incomplete records and records with missing key indicators, I retained 200 valid patents for empirical analysis. The sample includes both invention patents and utility model patents. Invention patents account for 182 records, while utility model patents account for 18 records. Because the number of utility model patents is relatively small, I did not separate the value effects of patent types in this study. Since actual transaction prices are rarely disclosed, I use the IncoPat value score as a proxy for patent value. This score is generated from patent big data analysis and reduces subjective judgment.

I discretize continuous indicators into five levels based on frequency distribution. This step is necessary for rough set processing and for reducing the influence of different measurement scales. I also apply min-max normalization to all data before neural network training. The normalization formula is:

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

Table 2 shows the discretization standard used in my study.

Indicator Level 1 Level 2 Level 3 Level 4 Level 5
C1 patent text quality [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 number of claims [1,5) [5,9) [9,13) [13,17) ≥17
C5 forward citations [0,2) [2,4) [4,6) [6,8) ≥8
C6 backward citations [0,2) [2,4) [4,6) [6,8) ≥8
C7 patent width [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 Valid — — —
C11 remaining life [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 Firm University — —
C15 patent family size [1,2) [2,3) [3,4) ≥4 —
C16 CPC classes [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

10. Attribute Reduction with Neighborhood Rough Sets

I treated the 18 indicators as condition attributes and the IncoPat value score as the decision attribute. Using MATLAB, I performed neighborhood rough set reduction. I tested different neighborhood radius parameters and significance thresholds. The parameter \(\lambda\) controls the neighborhood radius, and the significance threshold controls the minimum contribution of an attribute. Table 3 summarizes representative reduction results.

Significance Threshold λ Reduced Attribute Set
0.001 0.5 {1,2,3,4,7,8,10,11,12,13,14,15,16,17}
0.001 0.6 {1,2,3,4,5,6,7,8,11,12,14,16}
0.001 0.7 {1,2,3,4,5,6,7,8,10,11,12,14,16,17}
0.001 0.8 {1,2,3,4,5,6,7,8,10,11,12,14,16,17}
0.001 0.9 {1,2,3,5,6,7,8,10,11,12,13,14,16,17}
0.001 1.0 {1,2,3,4,5,6,7,8,12,13,14,16,17}
0.001 1.1 {1,3,4,7,8,17}
0.001 1.2 {1,3,4,7,8,17}
0.001 1.3 {1,4,7,8}
0.01 0.5 {3,4,8,11,12,15,16,17}
0.01 0.6 {1,2,3,4,5,6,7,8,11,12,14,16}
0.01 0.7 {1,2,3,4,5,7,8,11,12,16,17}
0.01 0.8 {1,2,3,4,5,7,8,11,12,16,17}
0.01 0.9 {1,2,3,5,6,7,8,10,11,12,13,14,16,17}
0.01 1.0 {1,3,4,5,6,7,8,13,14,16,17}
0.01 1.1 {1,3,4,7,8,17}
0.01 1.2 {1,3,4,7,8,17}
0.01 1.3 {1,4,7,8}

I then counted the frequency of each indicator across the reduction results. C8 appeared 18 times, C1 and C7 appeared 17 times, C3 and C4 appeared 16 times, C17 appeared 14 times, C16 appeared 12 times, C12 appeared 11 times, and C2, C5, and C11 appeared 10 times. I retained indicators that appeared more than ten times. The final reduced indicator set contains 11 indicators: C1, C2, C3, C4, C5, C7, C8, C11, C12, C16, and C17. Table 4 presents the final indicator system after reduction.

Primary Dimension Secondary Indicator
Text Value C1 patent text quality
Text Value C2 number of drawings
Text Value C3 number of inventors
Text Value C4 number of claims
Technical Value C5 forward citations
Technical Value C7 patent width
Technical Value C8 technical advancement
Legal Value C11 remaining life
Legal Value C12 pledge status
Market Value C16 CPC classes
Market Value C17 fee payment count

The reduction removes seven indicators: C6, C9, C10, C13, C14, C15, and C18. This does not mean these indicators are irrelevant in all contexts. Rather, within my sample and decision system, they are redundant given the retained indicators. Removing them reduces model complexity and improves computational efficiency. For the electric vehicle battery pack patent domain, this reduced set still captures text quality, technical breadth, technical advancement, legal life, pledge constraints, market coverage, and maintenance behavior.

11. GWO-BP-AdaBoost Model Construction

I built the prediction model in MATLAB. The model uses the 11 reduced indicators as input neurons and one output neuron for patent value. I set three hidden layers with 600, 800, and 1200 neurons respectively. The training function is trainlm. The maximum number of training epochs is 1000, the learning rate is 0.0005, and the error target is \(10^{-5}\). The grey wolf optimizer is used to optimize the initial weights and biases of each BP network. AdaBoost is then used to combine multiple optimized BP networks into a strong predictor.

Table 5 summarizes the main parameters of the model.

Parameter Setting
Input neurons 11
Hidden layers 3
Hidden layer neurons 600, 800, 1200
Output neurons 1
Training function trainlm
Maximum epochs 1000
Error target \(10^{-5}\)
Learning rate 0.0005
Optimizer Grey wolf optimizer
Ensemble method AdaBoost

The GWO-BP-AdaBoost model is trained on 140 patent samples, which represent 70% of the data. The remaining 60 patent samples, which represent 30% of the data, are used as the test set. The test set is not used during training, so it provides a realistic check of the model’s generalization ability. I use the coefficient of determination \(R^2\), mean absolute percentage error (MAPE), and relative error to evaluate performance.

The MAPE is calculated as:

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

The coefficient of determination is:

$$R^2 = 1 – \frac{\sum_{i=1}^n (y_i – \hat{y}_i)^2}{\sum_{i=1}^n (y_i – \bar{y})^2}$$

12. Model Training and Prediction Results

The convergence behavior of the GWO-BP-AdaBoost model shows that the fitness decreases until approximately the 46th iteration and then stabilizes. This suggests that the algorithm reaches a stable solution around the 46th iteration. Continuing beyond this point does not significantly improve performance. The training regression coefficient is 0.96531, which is close to 1. This indicates that the model fits the training data well and captures the nonlinear relationship between patent indicators and patent value.

On the test set, 91.67% of the patents have a relative error below 10%. This means that for most lithium-ion power battery patents, the predicted value is close to the actual value score. The remaining errors may arise from unobserved factors such as transaction timing, strategic value within a specific electric vehicle battery pack platform, or unrecorded licensing relationships. Nevertheless, the overall accuracy is strong enough to support screening and ranking tasks.

I compare three models: a standard BP neural network, a BP neural network with AdaBoost, and the proposed GWO-BP-AdaBoost model. Table 6 reports the comparison.

Model Accuracy within 10% Error MAPE \(R^2\) Main Observation
BPNN 63.00% Above 10% Lower Large errors on several patents; weak generalization
BPNN-AdaBoost Higher than BPNN Reduced Improved AdaBoost reduces overall error but some outliers remain
GWO-BP-AdaBoost 91.67% Below 10% 0.907 Best accuracy, lowest error, strongest fit

The standard BPNN model has 22 out of 60 test samples with relative errors greater than 0.1. Three patents have especially large relative errors, reaching 0.88, 1.38, and 2.59. This confirms that a single BP network can be unstable when the data are noisy and the value signal is weak. The BPNN-AdaBoost model reduces the number of large-error samples to 6, but one patent still has a relative error of 0.96. The GWO-BP-AdaBoost model further reduces the number of large-error samples to 5, and all relative errors are below 0.3. This is a substantial improvement. The MAPE of the proposed model is below 10%, and its \(R^2\) is 0.907, which is significantly higher than the other two models. The results support the effectiveness of combining neighborhood rough sets with GWO and AdaBoost for evaluating lithium-ion power battery patents.

13. Case-Based Evaluation

To test the practical applicability of the model, I selected five lithium-ion power battery patents from a leading manufacturer in the electric vehicle battery pack supply chain. I use a public market assumption and a base date of October 31, 2024. The value type is market value. The evaluation purpose is to provide a reference for transaction pricing and portfolio management. I retrieve patent information from public patent databases and transaction platforms. Table 7 presents the basic information of the five patents.

Patent Technology Focus Application Number
Patent A Soft-pack lithium-ion battery module and connection method CN201810408111.4
Patent B Lithium-ion battery current collector foil rolling and sizing device CN201810053598.9
Patent C Lithium-sulfur battery solid electrolyte membrane and preparation method CN202010326547.6
Patent D Lithium-ion battery negative electrode pre-lithiation device and method CN201710861791.0
Patent E Negative electrode sheet and lithium-ion battery containing the same CN202211672410.1

Table 8 shows the values of the 11 reduced indicators for the five patents.

Indicator Patent A Patent B Patent C Patent D Patent E
C1 patent text quality 10 9 9 10 15
C2 number of drawings 5 3 2 4 1
C3 number of inventors 5 3 3 4 5
C4 number of claims 7 8 14 10 10
C5 forward citations 1 5 0 4 3
C7 patent width 7 3 4 1 2
C8 technical advancement 10 5 9 10 8
C11 remaining life 13.55 13.27 15.53 12.9 18.2
C12 pledge status No No No No No
C16 CPC classes 3 4 9 2 1
C17 fee payment count 6 5 4 7 2

After discretization, the data are input into the trained GWO-BP-AdaBoost model. Table 9 presents the predicted values and actual value scores. The average absolute error is 4.46%, and all five patents have relative errors below 10%. This indicates that the model can produce useful estimates for patents related to the electric vehicle battery pack.

Patent Actual Value Predicted Value Relative Error
Patent A 10 9.273 7.3%
Patent B 8 8.310 3.9%
Patent C 9 9.634 7.0%
Patent D 9 9.108 1.2%
Patent E 9 8.737 2.9%

14. Interpretation of Results

The results show that the reduced indicator set retains strong predictive power. Text quality, drawing count, inventor count, claim count, forward citations, patent width, technical advancement, remaining life, pledge status, CPC classes, and fee payment count collectively explain a large share of patent value variation. This makes sense in the electric vehicle battery pack context. A patent with broad technical advancement and many CPC classes is more likely to cover multiple applications within an electric vehicle battery pack. A patent with a long remaining life and active fee payments signals that the owner expects future commercial benefit. A patent with more claims and drawings often has a more developed technical description, which can support licensing and litigation.

The GWO-BP-AdaBoost model outperforms the standard BPNN and BPNN-AdaBoost models because it combines global search with ensemble learning. The grey wolf optimizer finds better initial parameters and reduces the chance of local minima. AdaBoost then focuses on samples that are difficult to predict, which improves the overall accuracy. For patent value assessment, this is important because high-value patents are rare and may be outliers in the data. A standard BP network may underfit these outliers, while an AdaBoost ensemble can allocate more attention to them.

15. Managerial and Policy Implications

For firms that develop an electric vehicle battery pack, the proposed model can be used to screen patent portfolios. Managers can input patent indicators into the model and obtain a value estimate. This helps them decide which patents to maintain, which to license, and which to abandon. It also helps them identify patents that may be undervalued in internal accounting. For investors, the model can support due diligence by providing a quantitative signal of patent quality. For technology transfer offices, the model can help prioritize patents for commercialization. For policymakers, the model can support the design of innovation incentives by identifying which patent characteristics are associated with higher value in the electric vehicle battery pack ecosystem.

Because the model uses publicly available patent indicators, it can be applied at scale. It does not require confidential cost data or private transaction prices. This is a major advantage in the lithium-ion power battery field, where transaction data are scarce. The model can also be updated as new patents are granted and as new citations accumulate. This makes it suitable for dynamic monitoring of an electric vehicle battery pack patent portfolio.

16. Limitations and Future Research

My study has several limitations. First, the sample size is limited to 200 patents. Although this is sufficient for an initial model, a larger sample would improve generalization. Second, the value label is based on a patent database score rather than actual transaction prices. If more transaction data become available, the model can be retrained with actual prices. Third, some qualitative indicators are quantified through database scores, which may introduce measurement error. Future research can improve the quantification of qualitative indicators by using text mining and natural language processing. Fourth, the electric vehicle battery pack is a rapidly evolving system. New architectures such as cell-to-pack and cell-to-body may change the relative importance of patent indicators. The model should therefore be updated periodically.

Future research can also explore other optimization algorithms, such as particle swarm optimization, genetic algorithms, and differential evolution. It can compare different ensemble methods, such as random forests, gradient boosting, and stacked generalization. It can also apply deep learning models to patent text and citation networks. Another promising direction is to integrate technical standards, supply chain data, and battery performance data into the evaluation framework. This would allow a more comprehensive assessment of how a patent contributes to the electric vehicle battery pack.

17. Conclusion

In this study, I developed a hybrid framework for evaluating lithium-ion power battery patent value. I began with an 18-indicator system covering text value, technical value, legal value, and market value. I used neighborhood rough sets to reduce the indicator set to 11 indicators. I then built a GWO-BP-AdaBoost neural network to predict patent value. The model achieved 91.67% accuracy within a 10% relative error threshold on the test set, with a MAPE below 10% and an \(R^2\) of 0.907. In the case evaluation, the average relative error was 4.46%. These results indicate that the proposed method is effective for lithium-ion power battery patent value assessment.

The broader significance of this work is that it provides a data-driven, scalable, and objective approach to valuing patents that support the electric vehicle battery pack. As the electric vehicle industry continues to grow, the ability to identify and evaluate high-value patents will become increasingly important. My method contributes to that goal by combining rough set theory, neural networks, and ensemble optimization. It reduces subjective judgment, handles nonlinear relationships, and can be applied to large patent portfolios. I believe that further development of this framework will support better decisions in patent management, technology transfer, financing, and innovation policy for the electric vehicle battery pack.

Scroll to Top