Shared electric bicycle battery health prediction and off-peak charging optimization method based on big data

Abstract

The invention is suitable for the technical field of charging optimization, and provides a shared electric bicycle battery health prediction and off-peak charging optimization method based on big data, and the method comprises the steps: inputting obtained battery operation data into a pre-trained battery health prediction model, and outputting a health state prediction value of a battery; constructing a multi-target optimization model aiming at reducing the overall operation cost of the shared electric bicycle fleet, prolonging the overall service life of a battery pack of the fleet and stabilizing the regional power grid load; based on the determined health state prediction value, the time-of-use electricity price data, the power grid load data and the charging demand data, a genetic algorithm is adopted for solving, and a peak shifting charging scheduling scheme is obtained; and issuing the off-peak charging scheduling scheme to a charging control terminal of the shared electric bicycle operation platform so as to execute centralized off-peak charging operation. The method solves the problems of inaccurate prediction, uneconomical scheduling and disjunction of a power grid and actual operation in the prior art, and realizes collaborative optimization of operation efficiency, battery life and power grid friendliness.

Description

Technical Field

[0001] This invention relates to the field of charging optimization technology, and in particular to a method for predicting the battery health and optimizing off-peak charging of shared electric bicycles based on big data. Background Technology

[0002] As an important component of the sharing economy and green travel, shared electric bicycles have been widely used in urban short-distance transportation in recent years. Their large-scale operation relies heavily on the performance and availability of onboard batteries; therefore, accurate assessment of battery health and efficient charging management are key technological factors determining operating costs and user experience.

[0003] In existing technologies, battery health status prediction is mainly based on empirical models using laboratory data or simple statistical methods. For example, health is linearly estimated by monitoring changes in battery voltage, current, and internal resistance, combined with the number of charge-discharge cycles. However, these methods struggle to capture the nonlinear degradation characteristics of batteries under complex real-world operating conditions, leading to unreliable predictions and potentially causing premature battery failure or safety hazards during operation. On the other hand, in the field of charging scheduling, existing shared electric bicycle operation management systems mostly employ scheduling strategies based on fixed times or simple threshold rules. While these solutions guarantee a basic power supply to vehicles to some extent, they fail to comprehensively consider multi-dimensional information such as time-of-use pricing, real-time regional grid load, and individual battery health differences at the system level, resulting in suboptimal operational economics and resource utilization efficiency.

[0004] In view of this, a method for predicting the battery health of shared electric bicycles and optimizing off-peak charging based on big data is proposed. Summary of the Invention

[0005] This invention provides a big data-based method for predicting the health of shared electric bicycle batteries and optimizing off-peak charging. This method addresses the problem that the failure to comprehensively consider multi-dimensional information such as time-of-use pricing, real-time load of the regional power grid, and individual battery health differences at the system level leads to suboptimal operational economics and resource utilization efficiency.

[0006] This invention provides a method for predicting the battery health of shared electric bicycles and optimizing off-peak charging based on big data, including: Obtain battery operation data for multiple vehicles in the shared electric bicycle fleet, time-of-use electricity price data within the operating area, grid load data for the corresponding area, and charging demand data; The battery operating data is input into a pre-trained battery health prediction model, which outputs a predicted value of the battery's health status. The battery health prediction model is a fusion model based on a long short-term memory network and an attention mechanism, used to model the long-term dependency of battery performance degradation. A multi-objective optimization model is constructed with the goals of reducing the overall operating cost of the shared electric bicycle fleet, extending the overall lifespan of the fleet's battery packs, and easing the regional power grid load. Based on the predicted health status, the time-of-use electricity price data, the power grid load data, and the charging demand data, a genetic algorithm is used to solve the multi-objective optimization model to obtain a peak-shaving charging scheduling scheme. The off-peak charging scheduling scheme is sent to the charging control terminal of the shared electric bicycle operation platform. The off-peak charging scheduling scheme represents the charging status of each vehicle in the fleet in multiple time segments of a future preset period in the form of a scheduling matrix, and controls the corresponding charging facilities to perform centralized off-peak charging operations.

[0007] Furthermore, the battery health prediction model includes a sequentially connected multidimensional feature extraction layer, a bidirectional long short-term memory network layer, an environmental awareness attention layer, and a health status regression layer, specifically: The multidimensional feature extraction layer is used to extract a time-domain feature set containing statistical features, fluctuation features and trend features from the battery operation data, and to extract a frequency-domain feature set characterizing the changes in the internal chemical state of the battery through fast Fourier transform. The bidirectional long short-term memory network layer takes as input the spliced ​​time-domain feature sequence and frequency-domain feature sequence, and is used to model the long-term dependency of the battery operation data in the time dimension from both forward and backward directions, and outputs a hidden state sequence containing complete temporal context information. The environment-aware attention layer is used to receive the hidden state sequence and calculate the importance weight of the hidden state at each time step for the final health state prediction, generating a weighted context vector. The health status regression layer is used to map the context vector to the final battery health status prediction value.

[0008] Furthermore, the environment-aware attention layer is implemented through an attention computing network, specifically as follows: The input to the attention calculation network is the hidden state output by the bidirectional long short-term memory network layer; The attention computing network receives ambient temperature data from the battery operation data as additional prior conditions. The attention calculation network calculates the attention score using the following formula: in: For the first Attention score at each time step For the first The hidden state at each time step This is the embedded environmental temperature condition vector. , , For learnable parameter matrix, This is a bias term.

[0009] Furthermore, the training process of the battery health prediction model adopts a phased strategy based on consistency regularization, specifically including: In the first stage, an adaptive moment estimation optimizer is used to train all parameters of the model with a first learning rate. The first loss function used is the mean squared error between the predicted value and the true label, plus the second norm regularization term for all weight parameters of the model. In the second stage, the model parameters obtained from the first stage training are loaded, the parameters of the multidimensional feature extraction layer and the bidirectional long short-term memory network layer are frozen, and the parameters of the environment-aware attention layer and the health status regression layer are trained using a stochastic gradient descent optimizer at a second learning rate that is less than the first learning rate. In the second phase of training, the second loss function used is based on the first loss function, with the addition of a consistency regularization loss term; the consistency regularization loss term is obtained by calculating the variance of the predicted health status values ​​of the same vehicle under similar operating conditions and different ambient temperatures.

[0010] Furthermore, the reduction in the overall operating cost of the shared electric bicycle fleet is achieved by minimizing the economic cost of charging. To achieve this, the calculation formula is: in: The total number of vehicles in the shared electric bicycle fleet; This refers to the total number of time periods within the scheduling cycle, which is based on a 24-hour period. (The last part, "longer time periods," is a separate concept and doesn't need a direct translation.) It takes 1 hour; For vehicles The charging power; The binary decision variable represents the vehicle. During the period Whether it is charging or not, 1 indicates charging, 0 indicates not charging; For time period The time-of-use electricity pricing is based on the local time-of-use electricity pricing policy, which sets peak hours, average hours, and off-peak hours.

[0011] Furthermore, the extended overall lifespan of the fleet's battery packs is achieved by minimizing battery wear and tear costs. To achieve this, the calculation formula is: in: For vehicles The number of charging cycles within the scheduling period is included in the battery wear caused by frequency start-stop. For vehicles The average ambient temperature during charging is set based on local historical meteorological data; It is the battery activation energy, reflecting the temperature sensitivity of the battery's chemical reaction; This is the universal gas constant; For reference temperature, For vehicles Health status prediction value, , , The weighting coefficients are calibrated using battery cycle test data.

[0012] Furthermore, the load mitigation in the regional power grid is achieved by minimizing the peak load cost of the power grid. To achieve this, the calculation formula is: in: For the local power grid during the time period The base load is set based on historical load curves; This is the peak load penalty factor.

[0013] Furthermore, the constraints of the multi-objective optimization model include charging time window constraints based on vehicle usage patterns, battery thermal safety constraints considering high temperature and high humidity climates, power and spatial coupling constraints for centralized battery swapping cabinets, power guarantee constraints based on fleet regional scheduling needs, and soft constraints on total charging power during peak grid load periods.

[0014] Furthermore, the step of using a genetic algorithm to solve the multi-objective optimization model to obtain a staggered charging scheduling scheme includes: An initial population containing multiple individuals is randomly generated, with each individual representing a charging scheduling scheme. When generating the scheme, the charging operation completely avoids the morning and evening peak hours when local shared electric bicycles are prohibited, and the charging operation avoids the time when the local daytime ambient temperature exceeds the battery's safe charging threshold. The crossover and mutation probabilities are dynamically adjusted based on population diversity and the current generation number. Specifically, when population diversity is below a preset threshold, the crossover and mutation probabilities are increased; when population diversity is above a preset threshold, the crossover and mutation probabilities are decreased. As the generation number increases, the mutation probability exhibits an exponential decay trend. After crossover and mutation operations, the resulting new individuals undergo feasibility checks and repairs; the charging time for each vehicle remains continuous without interruption; the total charging power at the same charging station does not exceed the station's grid capacity limit; and the number of vehicles charging simultaneously at the same charging station does not exceed the physical slot limit.

[0015] Furthermore, the genetic algorithm also includes a local search enhancement step, which is executed at a preset cycle, specifically including: During the iteration that reaches the preset period, a certain proportion of individuals are randomly selected from the current population as candidate individuals to be optimized. For each candidate individual, multiple neighborhood scheduling schemes are generated through minor adjustments; the adjustment operations include: The charging start time of a vehicle can be shifted forward or backward during the local power grid's off-peak hours. The charging task of a vehicle is shifted from the off-peak hours of one charging station to the off-peak hours of another charging station; Evaluate the overall fitness of all neighborhood scheduling schemes; If there exists a neighborhood scheduling scheme with better fitness than the original candidate individual, then the original candidate individual in the population is replaced by the optimal neighborhood scheduling scheme. During the local search process, for vehicles located in key operating areas during the morning rush hour of the following day, priority will be given to ensuring that these vehicles reach the minimum battery level required for operation in the dispatch plan after the search.

[0016] As can be seen from the above technical solutions, the present invention has the following advantages: Based on multi-source data, this invention constructs a battery health prediction model that integrates long short-term memory networks and attention mechanisms to model the long-term dependency relationship of shared electric bicycle battery performance degradation, effectively improving the accuracy of health status prediction. Secondly, the predicted battery health value is used as input and fused with multi-source data such as time-of-use pricing and grid load to construct an optimization model aimed at reducing overall operating costs, extending battery life, and stabilizing regional grid load. An improved genetic algorithm is then used to solve this model, generating a scheduling scheme that considers regional characteristics and off-peak charging needs. Finally, through centralized control execution, this invention solves the problems of inaccurate prediction, uneconomical scheduling, and disconnect between the grid and actual operation in existing technologies, achieving synergistic optimization of operational efficiency, battery life, and grid friendliness. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of an embodiment of a method for predicting the health of shared electric bicycle batteries and optimizing off-peak charging based on big data in this invention. Figure 2 This is a schematic diagram of the architecture of the battery health prediction model in this invention; Figure 3 This is a flowchart illustrating the training process of the battery health prediction model in this invention. Figure 4 This is a schematic diagram of the improved genetic algorithm solution model in this invention. Detailed Implementation

[0018] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] Example 1 The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The method in this application will be described below from the perspective of system implementation. Please refer to... Figure 1 The method provided in this application includes the following steps: S1. Obtain battery operation data of multiple vehicles in the shared electric bicycle fleet, time-of-use electricity price data within the operating area, grid load data of the corresponding area, and charging demand data; In this embodiment, a battery management system deployed on the shared electric bicycle fleet collects real-time battery operation data for each vehicle. This data includes battery voltage, charging / discharging current, battery temperature, real-time state of charge, historical charging cycle count, and ambient temperature and vibration data collected by vehicle sensors. Simultaneously, time-of-use electricity price data is obtained from the public data interface of the power grid company in the operating area, including real-time electricity prices for peak, valley, and flat periods and their corresponding time periods. Power grid load data for the corresponding area is also obtained, including the base load curve, peak load periods, and the available capacity status of charging stations. Charging demand data is predicted and generated through comprehensive analysis of users' historical riding records, vehicle location data, riding trajectories, parking durations, and users' riding habits, usage frequency, and payment records.

[0020] After initial aggregation of the aforementioned multi-source heterogeneous data through the edge gateway, it is transmitted in real time with the Kafka message queue using the message queue telemetry transmission protocol. After an extraction-transformation-loading process, it is stored in a hybrid data storage layer consisting of a distributed file system, an HBase real-time database, a Radisson cache, and a MySQL relational database, providing a complete and consistent data foundation for subsequent processing.

[0021] S2. Input the battery operation data into the pre-trained battery health prediction model and output the predicted value of the battery health status; the battery health prediction model is a fusion model based on long short-term memory network and attention mechanism, used to model the long-term dependency of battery performance degradation; Please see Figure 2 The architecture diagram of the battery health prediction model is as follows: The battery health prediction model includes a sequentially connected multidimensional feature extraction layer, a bidirectional long short-term memory network layer, an environment-aware attention layer, and a health status regression layer. The multidimensional feature extraction layer is used to extract a set of time-domain features containing statistical features, fluctuation features and trend features from battery operation data, and to extract a set of frequency-domain features characterizing changes in the internal chemical state of the battery through fast Fourier transform. The bidirectional long short-term memory network layer takes as input the concatenated temporal and frequency-domain feature sequences, and is used to model the long-term dependencies of battery operation data in the time dimension from both forward and backward directions, and outputs a hidden state sequence containing complete temporal context information. An environment-aware attention layer is used to receive a sequence of hidden states and calculate the importance weight of the hidden state at each time step for the final health state prediction, generating a weighted context vector. A health status regression layer is used to map the context vector to the final battery health status prediction.

[0022] Specifically, the model architecture is a dedicated neural network that deeply integrates temporal processing, attention mechanisms, and domain knowledge. Its operation flow is as follows: 1. Input layer: The raw battery operating data input is a multivariate time series, including voltage, current, temperature, etc. 2. Multidimensional Feature Extraction Layer: This layer calculates and extracts time-domain features using a sliding window. These features describe the statistical, fluctuating, and trend characteristics of battery behavior over time. A Fast Fourier Transform (FFT) is applied to time-series signals, especially voltage and current, to convert the signals from the time domain to the frequency domain. Frequency-domain features reflect the characteristics and changes in the battery's internal electrochemical reactions, changes that are often difficult to detect in the time domain and are important indicators of early performance degradation. The extracted time-domain and frequency-domain feature sets are then concatenated along the feature dimension to form a comprehensive multidimensional fused feature sequence describing the battery's state.

[0023] 3. Bidirectional Long Short-Term Memory (LSTM) Network Layer: This layer aims to capture long-term temporal dependencies and bidirectional contextual information during battery degradation. The fused feature sequence obtained in the previous step is input into the forward LSTM and the backward LSTM. The forward LSTM processes the sequence from beginning to end, capturing positive causal dependencies; the backward LSTM processes the sequence from end to beginning, capturing the impact of future states on the present. The hidden states of the two LSTMs at each time step are combined to form the final hidden state sequence. Each time point in this sequence contains panoramic information about the entire sequence, providing a solid contextual foundation for accurate prediction.

[0024] 4. Environmental Awareness Attention Layer: The hidden state sequence output above, along with an additional environmental temperature condition vector, is input into a small neural network. The environmental temperature condition vector, i.e., the environmental temperature data, serves as additional prior knowledge. The attention score is calculated at each time step; the calculation formula is as follows: ,in: For the first Attention score at each time step For the first The hidden state at each time step This is the embedded environmental temperature condition vector. , , For learnable parameter matrix, This is a bias term.

[0025] Introducing temperature allows the model to dynamically adjust the importance of data at different time steps under varying temperature conditions. The attention score is normalized using the Softmax function to obtain the attention weight at each time step, with the sum of the weights being 1. The normalized attention weights are then multiplied element-wise by the corresponding hidden states and summed to generate a fixed-length context vector.

[0026] 5. Health Status Regression Layer: The above context vector is input into a regression network consisting of fully connected layers; after nonlinear transformation, the network outputs a continuous value, which is the final predicted value of battery health status.

[0027] Please see Figure 3 The training process of the battery health prediction model adopts a phased strategy based on consistency regularization, specifically including: In the first stage, an adaptive moment estimation optimizer is used to train all parameters of the model with a first learning rate. The first loss function used is the mean squared error between the predicted value and the true label, plus the second norm regularization term for all weight parameters of the model. In the second stage, the model parameters obtained from the first stage training are loaded, the parameters of the multidimensional feature extraction layer and the bidirectional long short-term memory network layer are frozen, and the parameters of the environment perception attention layer and the health status regression layer are trained using a stochastic gradient descent optimizer at a second learning rate that is less than the first learning rate. In the second phase of training, the second loss function used is based on the first loss function, with the addition of a consistency regularization loss term. The consistency regularization loss term is obtained by calculating the variance of the predicted health status values ​​of the same vehicle under similar operating conditions and different ambient temperatures.

[0028] Specifically, the training process is implemented in two stages. The first stage aims to enable all parameters of the model to quickly learn the basic characteristics and degradation patterns of battery data, finding a reasonably performing initial solution. The adaptive moment estimation optimizer here automatically adjusts the learning step size of each parameter based on historical gradient information, resulting in fast convergence and making it well-suited for large-scale exploration in the early stages. The first learning rate is a relatively large initial learning step size, allowing for significant updates to model parameters in the early training phase, thereby quickly reducing loss and approaching a better solution region. The first loss function = mean squared error + second norm regularization term; mean squared error measures the average squared difference between the model's predicted values ​​and the true labels; the second norm regularization term is L2 regularization, which penalizes excessively large weight values ​​in the model, forcing the model to learn simpler, more generalized features rather than over-relying on a few specific data points, preventing overfitting on the training data.

[0029] The first phase continues until the model's performance on the validation set no longer shows significant improvement. At this point, the model has developed good basic predictive capabilities. The output of this phase is a pre-trained model whose feature extractor and sequence processor have been well calibrated.

[0030] The second stage aims to improve the consistency of predictions across different ambient temperatures by fixing the backbone feature network, thereby obtaining a more robust final model. First, the optimal model parameters obtained in the first stage are loaded as the starting point for this stage, freezing the parameters of the multi-dimensional feature extraction layer and the bidirectional long short-term memory network layer. The optimizer is switched to a stochastic gradient descent optimizer, with a second learning rate that is a fine-grained step size smaller than the first learning rate. Since the backbone network parameters are frozen, only the last few layers need fine-tuning; the smaller learning rate ensures that the parameters are smoothly and accurately adjusted to their optimal positions. The second loss function equals the first loss function plus a consistency regularization loss term. The consistency regularization loss term is constructed by calculating the variance of the predicted health status values ​​of the same vehicle under similar battery state of charge and load current operating conditions at different ambient temperature periods. This loss term penalizes the model for giving different predictions for the same health status due to different temperatures, forcing the model to learn more fundamental battery degradation characteristics independent of temperature. This directly addresses the ambient temperature fluctuation problem faced by shared electric bicycles in the real world, greatly improving the model's practicality and robustness.

[0031] The second stage only updates the layers that are not frozen, namely the environment-aware attention layer and the health status regression layer; training continues until the loss function converges or the predetermined number of iterations is reached. The output of this stage is the final battery health prediction model that can be used in the production environment.

[0032] Based on the above description of the model architecture and training process, the real-time battery operation data collected from shared electric bicycles is input into the trained battery health prediction model. First, the multi-dimensional feature extraction layer automatically parses and fuses the time-domain statistical features of voltage and current sequences with the frequency-domain features representing the internal chemical state. Then, the bidirectional long short-term memory network layer captures the long-term temporal patterns of battery performance degradation from the fused features and outputs a hidden state sequence containing complete contextual information. The environmental awareness attention layer dynamically calculates the contribution weight of the hidden state at different times to the current prediction and generates a context vector of key degradation nodes by combining environmental temperature data. Finally, the health state regression layer maps this vector to the predicted value of battery health state.

[0033] S3. Construct a multi-objective optimization model with the goals of reducing the overall operating cost of the shared electric bicycle fleet, extending the overall lifespan of the fleet's battery packs, and easing the regional power grid load; based on the health status prediction value, time-of-use electricity price data, power grid load data, and charging demand data, use a genetic algorithm to solve the multi-objective optimization model and obtain a peak-shaving charging scheduling scheme; In this embodiment, the multi-objective optimization model is constructed based on the weighted summation method, integrating multiple conflicting objectives into a comprehensive objective function, and balancing the relative importance of each objective through weight coefficients. Simultaneously, the model fully considers the regional characteristics of shared electric bicycle operation and the demand for off-peak charging, setting a targeted constraint system to ensure that the generated scheduling scheme is both economical and efficient, and conforms to the actual operating scenario. The specific terms and constraints of the objective function are as follows: Reduce the overall operating cost of shared electric bicycle fleets by minimizing charging costs. To achieve this, the calculation formula is: in: The total number of vehicles in the shared electric bicycle fleet; This refers to the total number of time periods within the scheduling cycle, which is based on a 24-hour period. (The last part, "longer time periods," is a separate concept and doesn't need a direct translation.) It takes 1 hour; For vehicles The charging power; The binary decision variable represents the vehicle. During the period Whether it is charging or not, 1 indicates charging, 0 indicates not charging; For time period The time-of-use electricity pricing is based on the local time-of-use electricity pricing policy, which sets peak hours, average hours, and off-peak hours.

[0034] Extending the overall lifespan of a fleet's battery packs by minimizing battery degradation costs To achieve this, the calculation formula is: in: For vehicles The number of charging cycles within the scheduling period is included in the battery wear caused by frequency start-stop. For vehicles The average ambient temperature during charging is set based on local historical meteorological data; It is the battery activation energy, reflecting the temperature sensitivity of the battery's chemical reaction; This is the universal gas constant; For reference temperature, For vehicles Health status prediction value, , , The weighting coefficients are calibrated using battery cycle test data.

[0035] Smoothing regional grid load by minimizing peak grid load costs To achieve this, the calculation formula is: in: For the local power grid during the time period The base load is set based on historical load curves; This is the peak load penalty factor.

[0036] The constraints of the multi-objective optimization model include charging time window constraints based on vehicle usage patterns, battery thermal safety constraints considering high-temperature and high-humidity climates, power and spatial coupling constraints for centralized battery swapping stations, power guarantee constraints based on fleet regional scheduling needs, and soft constraints on total charging power during peak grid load periods. The expressions are as follows: Charging time window constraints based on vehicle usage patterns: in: The set of time periods during which charging is prohibited is set based on local traffic flow data; for example... This refers to the peak hours for vehicle use in the morning and evening. This time period can be adjusted according to the specific local traffic flow characteristics, forcibly locking the charging time to the overlap of the off-peak hours of vehicle demand and the off-peak hours of the power grid, thus achieving peak-shifting of vehicle use and charging.

[0037] Considering battery thermal safety constraints in high-temperature and high-humidity climates: in: For the local area Forecast ambient temperature for the period The maximum ambient temperature threshold for battery charging is set according to the battery's technical specifications. Local weather forecast data is used as a constraint input here to adapt to regional climate characteristics.

[0038] Power and space coupling constraints for centralized battery swapping cabinets: in: To belong to the A collection of vehicles at a centralized charging station or battery swapping station. For the first The total power limit of each site is determined by the capacity of the local power grid infrastructure. For the first The total number of physical slots for each battery swapping station; this constraint is for the centralized battery swapping station business model and takes into account the grid capacity at the site level.

[0039] Power supply guarantee constraints based on fleet area dispatching needs: in: This is for vehicles located in key areas to ensure smooth traffic flow during the morning rush hour the following day. For vehicles The minimum required battery charge for the area is set based on operational needs. This constraint deeply integrates regional vehicle scheduling with charging optimization, ensuring that the optimization results not only reduce costs but also directly improve the quality of operational services during the following morning rush hour.

[0040] Soft constraint on total charging power during peak grid load: in: This is the peak load period for the local power grid. for The maximum total charging power allowed during a given time period, which can be set in consultation with the grid operator.

[0041] Based on the above objective function and constraints, an improved genetic algorithm is used to solve the problem. Please refer to [link / reference]. Figure 4 The process mainly includes the following steps: 1. Randomly generate an initial population containing multiple individuals, each representing a charging scheduling scheme; when generating the scheme, the charging operation completely avoids the morning and evening peak hours when local shared electric bicycles are prohibited, and the charging operation avoids the time when the local daytime ambient temperature exceeds the battery's safe charging threshold. 2. Based on population diversity and the current iteration generation, dynamically adjust the crossover probability and mutation probability; specifically, when the population diversity is below a preset threshold, increase the crossover probability and mutation probability; when the population diversity is above the preset threshold, decrease the crossover probability and mutation probability; as the iteration generation increases, the mutation probability exhibits an exponential decay trend. 3. After crossover and mutation operations, the resulting new individuals are subject to feasibility checks and repairs; the charging time for each vehicle remains continuous without interruption; the total charging power at the same charging station does not exceed the upper limit of the station's grid capacity; and the number of vehicles charging simultaneously at the same charging station does not exceed the upper limit of the number of physical slots.

[0042] Specifically, in the population initialization phase, the two core rules of avoiding peak usage times in the morning and evening and avoiding high-temperature charging are directly embedded as hard constraints into the solution generation process, thus ensuring that all individuals are feasible solutions that meet actual operational needs from the outset. Secondly, in the adaptive parameter adjustment phase, the algorithm dynamically allocates exploration and development resources by monitoring population diversity in real time: when the population tends to be homogeneous, it actively increases the crossover and mutation probabilities, injecting new genes to prevent premature convergence to pseudo-peak-shifting schemes with local optima; when the population is too dispersed, it appropriately reduces the probabilities to promote the integration and convergence of superior genes. Simultaneously, the mutation probability decays exponentially with each generation, achieving a natural transition from extensively searching various potential peak-shifting periods in the early stages to fine-tuning charging times in the later stages. Finally, in the feasibility repair phase, the algorithm intelligently repairs invalid solutions that may be generated by genetic operations, addressing the unique charging continuity and site resource bottleneck issues in the centralized charging mode of shared electric bicycles. This ensures that each generation of the population strictly adheres to operational physical constraints, keeping the optimization process always on the correct feasible domain track.

[0043] In this embodiment, the genetic algorithm also inherits a local search enhancement step, which is executed according to a preset period, specifically including: 1. During the iteration that reaches the preset period, a certain proportion of individuals are randomly selected from the current population as candidate individuals to be optimized; 2. For each candidate individual, multiple neighborhood scheduling schemes are generated through minor adjustments; the adjustment operations include: 3. Shift the charging start time of a vehicle earlier or later during the local power grid's off-peak hours; 4. Adjust the charging task of a vehicle from the off-peak hours of one charging station to the off-peak hours of another charging station; 5. Evaluate the overall fitness of all neighborhood scheduling schemes; 6. If there exists a neighborhood scheduling scheme with better fitness than the original candidate individual, then the original candidate individual in the population is replaced by the optimal neighborhood scheduling scheme; 7. During the local search process, for vehicles located in key operating areas during the morning rush hour of the following day, priority should be given to ensuring that these vehicles reach the minimum battery level required for operation in the dispatch plan after the search.

[0044] Specifically, once the algorithm enters a preset local search cycle, it selects a subset of candidate individuals from the current population for in-depth optimization. Subsequently, for each candidate individual's represented scheduling scheme, within the critical time window of grid off-peak hours, it fine-tunes the charging start time of a particular vehicle. This aims to further leverage the benefits of lower electricity prices and load without impacting user experience. The algorithm also transfers vehicle charging tasks between the idle capacity of different charging stations to balance the load on each station and alleviate local resource bottlenecks. After generating a series of neighborhood schemes, the algorithm evaluates the overall fitness of each new scheme—the total cost after balancing electricity costs, battery degradation, and grid load—and replaces the original scheme with a better one, following a survival-of-the-fittest principle. Crucially, throughout the local search process, the algorithm prioritizes vehicles located in key operational areas during the following morning's peak hours, ensuring they still meet the minimum power requirements after optimization. This demonstrates the algorithm's absolute obedience to and support for higher-level operational strategies. This series of operations allows the algorithm not only to find globally optimal off-peak intervals but also to conduct in-depth analysis to find the most precise and economical charging scheduling points within the region.

[0045] S4. The off-peak charging scheduling plan is sent to the charging control terminal of the shared electric bicycle operation platform. The off-peak charging scheduling plan represents the charging status of each vehicle in the fleet in multiple time segments of a future preset period in the form of a scheduling matrix, and controls the corresponding charging facilities to perform centralized off-peak charging operations.

[0046] In this embodiment, the generated off-peak charging scheduling scheme is distributed to the charging control terminal of the shared electric bicycle operation platform in the form of a digital scheduling matrix via an encrypted communication network. The scheduling matrix is ​​a two-dimensional data structure, where rows represent each specific vehicle in the fleet, columns represent continuous time segments divided into a preset period, and the element values ​​in the matrix are binary charging status instructions, with 1 indicating charging and 0 indicating standby.

[0047] After receiving and parsing the scheduling matrix, the charging control terminal converts it into a series of executable commands with precise timing. Its core operation is as follows: the system automatically matches the corresponding time segment column in the scheduling matrix based on the current time, reads the instructions of all vehicles requiring charging within that time period, and then sends an activation command to the centralized charging cabinet or dedicated charging pile where the target vehicle is located via IoT protocol. Key execution logic includes: 1. Charging facilities only supply power to authorized vehicles during the time segments specified by the matrix, ensuring that the charging load is accurately distributed during the preset off-peak periods; 2. During the charging process, the charging facility monitors the charging status and battery parameters in real time and feeds the results back to the operation platform; 3. At the start of the planned charging period, if the system detects that a vehicle is in use or has been reserved by a user through GPS location, it will automatically mark the charging task of that vehicle as delayed. Based on the scheduling matrix and real-time grid load, the system will dynamically allocate a new charging window for the vehicle during the subsequent available off-peak hours, thereby maximizing the overall benefits of off-peak charging while ensuring the user's driving experience.

[0048] The above process realizes end-to-end automated control from optimization decision-making to physical execution, enabling the charging behavior of thousands of shared electric bicycles to act as a unified and schedulable load, accurately responding to grid signals and operational needs, and truly achieving large-scale, low-cost, grid-friendly centralized off-peak charging operation.

[0049] It is understood that those skilled in the art can combine various implementation methods in the above embodiments under the guidance of the above examples to obtain technical solutions with multiple implementation methods.

[0050] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the battery health and optimizing off-peak charging of shared electric bicycles based on big data, characterized in that, include: Obtain battery operation data for multiple vehicles in the shared electric bicycle fleet, time-of-use electricity price data within the operating area, grid load data for the corresponding area, and charging demand data; The battery operating data is input into a pre-trained battery health prediction model, which outputs a predicted value of the battery's health status. The battery health prediction model is a fusion model based on a long short-term memory network and an attention mechanism, used to model the long-term dependency of battery performance degradation. A multi-objective optimization model is constructed with the goals of reducing the overall operating cost of the shared electric bicycle fleet, extending the overall lifespan of the fleet's battery packs, and easing the regional power grid load. Based on the predicted health status, the time-of-use electricity price data, the power grid load data, and the charging demand data, a genetic algorithm is used to solve the multi-objective optimization model to obtain a peak-shaving charging scheduling scheme. The off-peak charging scheduling scheme is sent to the charging control terminal of the shared electric bicycle operation platform. The off-peak charging scheduling scheme represents the charging status of each vehicle in the fleet in multiple time segments of a future preset period in the form of a scheduling matrix, and controls the corresponding charging facilities to perform centralized off-peak charging operations.

2. The method for predicting the battery health of shared electric bicycles and optimizing off-peak charging based on big data as described in claim 1, characterized in that, The battery health prediction model includes a sequentially connected multidimensional feature extraction layer, a bidirectional long short-term memory network layer, an environmental awareness attention layer, and a health status regression layer, specifically: The multidimensional feature extraction layer is used to extract a time-domain feature set containing statistical features, fluctuation features and trend features from the battery operation data, and to extract a frequency-domain feature set characterizing the changes in the internal chemical state of the battery through fast Fourier transform. The bidirectional long short-term memory network layer takes as input the spliced ​​time-domain feature sequence and frequency-domain feature sequence, and is used to model the long-term dependency of the battery operation data in the time dimension from both forward and backward directions, and outputs a hidden state sequence containing complete temporal context information. The environment-aware attention layer is used to receive the hidden state sequence and calculate the importance weight of the hidden state at each time step for the final health state prediction, generating a weighted context vector. The health status regression layer is used to map the context vector to the final battery health status prediction value.

3. The method for predicting the battery health of shared electric bicycles and optimizing off-peak charging based on big data as described in claim 2, characterized in that, The environment-aware attention layer is implemented through an attention computation network, specifically as follows: The input to the attention calculation network is the hidden state output by the bidirectional long short-term memory network layer; The attention computing network receives ambient temperature data from the battery operation data as additional prior conditions. The attention calculation network calculates the attention score using the following formula: in: For the first Attention score at each time step For the first The hidden state at each time step. This is the embedded environmental temperature condition vector. , , For learnable parameter matrix, This is a bias term.

4. The method for predicting the battery health and optimizing off-peak charging of shared electric bicycles based on big data as described in claim 3, characterized in that, Its features are, The training process of the battery health prediction model adopts a phased strategy based on consistency regularization, specifically including: In the first stage, an adaptive moment estimation optimizer is used to train all parameters of the model with a first learning rate. The first loss function used is the mean squared error between the predicted value and the true label, plus the second norm regularization term for all weight parameters of the model. In the second stage, the model parameters obtained from the first stage training are loaded, the parameters of the multidimensional feature extraction layer and the bidirectional long short-term memory network layer are frozen, and the parameters of the environment-aware attention layer and the health status regression layer are trained using a stochastic gradient descent optimizer at a second learning rate that is less than the first learning rate. In the second phase of training, the second loss function used is based on the first loss function, with the addition of a consistency regularization loss term; the consistency regularization loss term is obtained by calculating the variance of the predicted health status values ​​of the same vehicle under similar operating conditions and different ambient temperatures.

5. The method for predicting the battery health of shared electric bicycles and optimizing off-peak charging based on big data as described in claim 1, characterized in that, The reduction of overall operating costs for shared electric bicycle fleets is achieved by minimizing charging costs. To achieve this, the calculation formula is: in: The total number of vehicles in the shared electric bicycle fleet; This refers to the total number of time periods within the scheduling cycle, which is based on a 24-hour period. (The last part, "longer time periods," is a separate concept and doesn't need a direct translation.) It takes 1 hour; For vehicles The charging power; The binary decision variable represents the vehicle. During the period Whether it is charging or not, 1 indicates charging, 0 indicates not charging; For time period The time-of-use electricity pricing is based on the local time-of-use electricity pricing policy, which sets peak hours, average hours, and off-peak hours.

6. The method for predicting the battery health of shared electric bicycles and optimizing off-peak charging based on big data as described in claim 1, characterized in that, The extension of the overall lifespan of the fleet's battery packs is achieved by minimizing battery wear and tear costs. To achieve this, the calculation formula is: in: For vehicles The number of charging cycles within the scheduling period is included in the battery wear caused by frequency start-stop. For vehicles The average ambient temperature during charging is set based on local historical meteorological data; It is the battery activation energy, reflecting the temperature sensitivity of the battery's chemical reaction; This is the universal gas constant; For reference temperature, For vehicles Health status prediction value , , The weighting coefficients are calibrated using battery cycle test data.

7. The method for predicting the battery health and optimizing off-peak charging of shared electric bicycles based on big data as described in claim 1, characterized in that, The method of mitigating regional grid load is achieved by minimizing the cost of peak grid load. To achieve this, the calculation formula is: in: For the local power grid during the time period The base load is set based on historical load curves; This is the peak load penalty factor.

8. The method for predicting the battery health and optimizing off-peak charging of shared electric bicycles based on big data according to claim 1, characterized in that, The constraints of the multi-objective optimization model include charging time window constraints based on vehicle usage patterns, battery thermal safety constraints considering high temperature and humidity climates, power and spatial coupling constraints for centralized battery swapping cabinets, power guarantee constraints based on fleet regional scheduling needs, and soft constraints on total charging power during peak grid load periods.

9. The method for predicting the battery health and optimizing off-peak charging of shared electric bicycles based on big data as described in claim 1, characterized in that, The method of solving the multi-objective optimization model using a genetic algorithm to obtain a staggered charging scheduling scheme includes: An initial population containing multiple individuals is randomly generated, with each individual representing a charging scheduling scheme. When generating the scheme, the charging operation completely avoids the morning and evening peak hours when local shared electric bicycles are prohibited, and the charging operation avoids the time when the local daytime ambient temperature exceeds the battery's safe charging threshold. The crossover and mutation probabilities are dynamically adjusted based on population diversity and the current generation number. Specifically, when population diversity is below a preset threshold, the crossover and mutation probabilities are increased; when population diversity is above a preset threshold, the crossover and mutation probabilities are decreased. As the generation number increases, the mutation probability exhibits an exponential decay trend. After crossover and mutation operations, the resulting new individuals undergo feasibility checks and repairs; the charging time for each vehicle remains continuous without interruption; the total charging power at the same charging station does not exceed the station's grid capacity limit; and the number of vehicles charging simultaneously at the same charging station does not exceed the physical slot limit.

10. The method for predicting the battery health and optimizing off-peak charging of shared electric bicycles based on big data according to claim 9, characterized in that, The genetic algorithm also inherits a local search enhancement step, which is executed at a preset cycle, specifically including: During the iteration that reaches the preset period, a certain proportion of individuals are randomly selected from the current population as candidate individuals to be optimized. For each candidate individual, multiple neighborhood scheduling schemes are generated through minor adjustments; the adjustment operations include: The charging start time of a vehicle can be shifted forward or backward during the local power grid's off-peak hours. The charging task of a vehicle is shifted from the off-peak hours of one charging station to the off-peak hours of another charging station; Evaluate the overall fitness of all neighborhood scheduling schemes; If there exists a neighborhood scheduling scheme with better fitness than the original candidate individual, then the original candidate individual in the population is replaced by the optimal neighborhood scheduling scheme. During the local search process, for vehicles located in key operating areas during the morning rush hour of the following day, priority will be given to ensuring that these vehicles reach the minimum battery level required for operation in the dispatch plan after the search.

Images