A method, device and medium for energy management of hybrid vehicles based on green wave traffic.

Abstract

This specification discloses a method, device, and medium for hybrid vehicle energy management based on green wave traffic, relating to the field of hybrid vehicle energy management technology. The method includes: acquiring the current motion state data of the hybrid vehicle and obtaining the traffic light phase timing information of the current intersection through a vehicle-to-infrastructure communication module to determine the guiding speed feature vector; extracting the information reception delay from the traffic light phase timing information to determine the multimodal observation state vector; inputting the multimodal observation state vector into a probabilistic action network in a pre-trained soft actor-critic network, outputting mean parameters and standard deviation parameters, calculating the reliability coefficient based on the standard deviation parameter, and using the reliability coefficient to generate a target equivalence factor; constructing an evaluation index function for instantaneous equivalent energy consumption, and solving for the engine target torque command and motor target torque command that minimize the evaluation index function within the feasible set of powertrain physical constraints, thereby realizing hybrid vehicle energy management.

Description

Technical Field

[0001] This specification relates to the field of hybrid vehicle energy management technology, and in particular to a hybrid vehicle energy management method, device and medium based on green wave traffic. Background Technology

[0002] With the rapid development of intelligent transportation systems and vehicle-to-infrastructure (V2I) communication technologies, hybrid electric vehicles (HEVs) can obtain real-time traffic light phase and timing information at intersections ahead through V2I modules. Based on this, green wave traffic flow is gradually becoming a key application scenario for improving urban traffic efficiency and reducing frequent vehicle starts and stops. Due to their multi-power source coupling and energy recovery mechanisms, HEVs possess significant energy-saving potential under urban green wave conditions. Therefore, how to optimize the energy management strategy of HEVs using V2I information has become a research hotspot in this field.

[0003] In existing technologies, various energy management methods for hybrid electric vehicles have been proposed. For example, one type of method employs rule-based control logic or fuzzy control strategies. These methods offer good engineering robustness and high real-time performance, but the control rules are pre-fixed and cannot adaptively adjust the power distribution between the engine and electric motor according to dynamically changing green wave windows, resulting in limited improvements in fuel economy. Another type of method employs optimization-based strategies, such as the equivalent fuel consumption minimization strategy and its adaptive form, which balance fuel and electric energy consumption through an equivalent factor. However, the equivalent factor in this type of method is typically a ratio. integral Differential controllers adjust based on battery state-of-charge (SOC) deviations, but this adjustment is inherently lagging. In green wave scenarios, when vehicle speed boundaries change drastically, this can easily cause oscillations in the equivalent factor, leading to increased SOC fluctuations and hindering further fuel economy improvements. Another approach uses deep reinforcement learning to directly output torque commands for the engine and motor. While this method can handle high-dimensional nonlinear data, it is prone to control instability in the absence of constraint mechanisms. Furthermore, most existing reinforcement learning methods rely solely on the vehicle's own state, failing to fully utilize future speed guidance information provided by vehicle-to-infrastructure (V2I) communication and data loss uncertainties in the V2I environment, resulting in insufficient robustness in complex dynamic traffic scenarios.

[0004] In summary, when using vehicle-road cooperative information for energy management of hybrid electric vehicles during green wave traffic, the uncertainty of the communication environment cannot be dynamically perceived, and the equivalent factor adjustment method lags behind changes in vehicle speed boundaries. This results in insufficient adaptability of the energy management strategy to complex green wave dynamic environments and poor decision-making stability under abnormal communication conditions, making it difficult to balance the fuel economy and system robustness of hybrid electric vehicles in green wave traffic scenarios. Summary of the Invention

[0005] This specification provides one or more embodiments of a hybrid vehicle energy management method, device, and medium based on green wave traffic, to solve the following technical problems: In the process of using vehicle-road cooperative information for hybrid vehicle green wave traffic energy management, it is impossible to dynamically perceive the uncertainty of the communication environment and the equivalent factor adjustment method lags behind the changes in vehicle speed boundaries. There are problems such as insufficient adaptability of energy management strategies to complex green wave dynamic environments and poor decision stability under abnormal communication conditions, making it difficult to balance the fuel economy and system robustness of hybrid vehicles in green wave traffic scenarios.

[0006] One or more embodiments of this specification employ the following technical solutions: This specification provides one or more embodiments of a hybrid vehicle energy management method based on green wave traffic. The method includes: acquiring the current motion state data of the hybrid vehicle and acquiring the traffic light phase timing information of the current intersection through a vehicle-to-infrastructure communication module; determining a guiding speed feature vector to characterize future speed trends based on the traffic light phase timing information and the current vehicle speed in the current motion state data; extracting information reception delay from the traffic light phase timing information; determining a multimodal observation state vector based on the current motion state data, the guiding speed feature vector, and the information reception delay; and inputting the multimodal observation state vector into... The probabilistic action network in the pre-trained soft actor-critic network outputs mean and standard deviation parameters to describe the probability distribution of the equivalent factor. A reliability coefficient is calculated based on the standard deviation parameter, and this coefficient is used to weight and fuse the preset dual-strategy equivalent factors to generate a target equivalent factor. Based on the current vehicle speed, current driving torque demand, and the target equivalent factor, an evaluation index function for instantaneous equivalent energy consumption is constructed. Within a preset feasible set of powertrain physical constraints, the engine target torque command and motor target torque command that minimize the evaluation index function are solved to achieve hybrid vehicle energy management.

[0007] This specification provides one or more embodiments of a hybrid vehicle energy management device based on green wave traffic, including: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the above-described method.

[0008] This specification provides one or more embodiments of a non-volatile computer storage medium storing computer-executable instructions configured to perform the above-described method.

[0009] The above-described at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: By extracting the guiding speed feature vector through rolling time-domain extrapolation, the spatiotemporal constraints of traffic lights at intersections are transformed into a quantitative representation of future speed trends. This enables the energy management strategy to be predictive, proactively adjusting the battery state of charge before the green light window arrives, reducing engine intervention in inefficient operating ranges, and thus significantly improving fuel economy. Utilizing the probability distribution of the equivalent factors output by a soft actor-critic network, environmental uncertainty is quantified in real-time using the standard deviation parameter. Based on this, a confidence coefficient is calculated, achieving adaptive fusion decision-making that fully exploits the agent's optimization potential under high confidence and automatically degenerates into a conservative state of charge maintenance strategy under high uncertainty. This significantly enhances the system's robustness and safety in complex real-world environments. A hierarchical architecture is adopted, using the equivalent factors output by the agent as weight parameters for the lower-level equivalent fuel consumption minimization strategy. The optimal torque combination is solved within a feasible set of strict powertrain physical constraints, preserving the global optimization capability of the data-driven strategy while ensuring the physical realizability and engineering safety of each set of instructions. This approach continuously reduces energy consumption while ensuring driving safety and component reliability, making it suitable for hybrid vehicles in green wave traffic scenarios. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart illustrating a hybrid vehicle energy management method based on green wave traffic, provided as an embodiment of this specification; Figure 2 A schematic diagram of a rolling time-domain dynamic velocity boundary pre-simulation provided for an embodiment of this specification; Figure 3 A schematic diagram of traffic light phase constraints and vehicle spatiotemporal trajectory provided in the embodiments of this specification; Figure 4 A schematic diagram of the system architecture of a hybrid vehicle energy management system based on green wave traffic, provided for embodiments of this specification; Figure 5 This is a schematic diagram of the overall process of a layered strategy provided in the embodiments of this specification; Figure 6 This is a schematic diagram illustrating the SOC maintenance effect provided in the embodiments of this specification; Figure 7This is a schematic diagram of the engine operating point in a fuel economy scenario provided in an embodiment of this specification. Figure 8 This is a schematic diagram of the structure of a hybrid vehicle energy management device based on green wave traffic, provided as an embodiment of this specification. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0012] This specification provides an embodiment of a hybrid vehicle energy management method based on green wave traffic. It should be noted that the execution entity in this specification embodiment can be a server or any device with data processing capabilities. Figure 1 This is a flowchart illustrating a hybrid vehicle energy management method provided in an embodiment of this specification, as shown below. Figure 1 As shown, the main steps include the following: Step S101: Obtain the current motion state data of the hybrid vehicle, and obtain the traffic light phase timing information of the current intersection through the vehicle-road communication module, so as to determine the guiding speed feature vector to characterize the future speed trend based on the traffic light phase timing information and the current vehicle speed in the current motion state data.

[0013] In one embodiment of this specification, after the hybrid vehicle starts and enters the road, at the beginning of a control cycle, the current motion state data of the hybrid vehicle is first collected from the vehicle sensor network via the vehicle controller local area network bus or the vehicle Ethernet. The current motion state data includes the current vehicle speed collected by the vehicle speed sensor, the current battery state of charge reported by the battery management system, and the current drive torque demand calculated by the vehicle controller based on the accelerator pedal opening. The current drive torque demand reflects the driver's expectation of the vehicle's driving force at the current moment.

[0014] Simultaneously, the vehicle-to-infrastructure (V2I) communication module receives the current intersection traffic light phase and timing information. The V2I module establishes a connection with the roadside equipment using dedicated short-range communication technology or cellular vehicle-to-everything (V2X) communication technology. The onboard controller sends a request signal containing the vehicle's identification and current location to the roadside equipment. In response, the roadside equipment transmits back the current traffic light status (red light, green light) and remaining duration of each phase at the intersection as traffic light phase and timing information. The information reception delay is calculated based on the received time and the timestamp carried in the information. Traffic light phase and timing information, also known as traffic light phase timing information, indicates the right-of-way status of the nearest intersection ahead of the vehicle and the corresponding time window boundaries, providing constraints for subsequent vehicle speed guidance.

[0015] Based on the traffic light phase timing information and the current vehicle speed in the current motion state data, a guiding speed feature vector is determined to characterize the future speed trend. Specifically, this includes: predicting the feasible upper and lower speed limits of the hybrid vehicle at each discrete time point within a preset future time period based on the traffic light phase timing information and the current vehicle speed, and determining a sequence of feasible speed domains; performing an arithmetic average of the feasible upper and lower speed limits at each discrete time point to obtain a guiding speed reference value for that discrete time point; and arranging multiple guiding speed reference values ​​in chronological order and then downsampling to determine the guiding speed feature vector.

[0016] Based on the traffic light phase timing information and the current vehicle speed, predict the feasible upper and lower speed limits of the hybrid vehicle at each discrete time point within a preset future duration, and determine the speed feasible domain sequence. Specifically, this includes: starting from the current sampling time, iterating backward through discrete time points with a preset time step until the preset future duration is reached; at each discrete time point, determining whether the vehicle's virtual position crosses the stop line of the traffic light ahead based on the traffic light phase timing information; if not, then based on the green light start and green light end times in the current traffic light phase timing information, combined with the hybrid vehicle's... The remaining distance between the current position and the stop line of the traffic light ahead is used to calculate the upper and lower feasible speed limits, respectively. The upper feasible speed limit is the fastest speed at which the hybrid vehicle can reach the stop line after the green light turns on, and the lower feasible speed limit is the slowest speed at which the hybrid vehicle can reach the stop line before the green light turns off. If the stop line has already been crossed, a safe guiding speed is calculated based on the preset maximum comfort deceleration to stop the hybrid vehicle before the stop line without crossing the stop line of the traffic light ahead, and both the upper and lower feasible speed limits are set to the safe guiding speed.

[0017] In one embodiment of this specification, the guiding speed feature vector for characterizing future speed trends is determined based on the traffic light phase timing information and the current vehicle speed in the current motion state data, specifically in the following manner: First, a rolling prediction model based on physical spatiotemporal constraints is constructed, taking the current sampling time as the starting point. Starting from the upstream section, the forward duration is calculated. Time step Evolutionary predictions, such as Figure 2 As shown, Figure 2 This specification provides a schematic diagram of a rolling time-domain dynamic velocity boundary pre-simulation for an embodiment. Within the prediction time bound, the kinematic characteristics of the vehicle in the future are simulated, and the virtual position is iterated in real time. With virtual speed : ,in, The simulated acceleration is generated based on the target vehicle speed feedback. The target vehicle speed refers to the desired vehicle speed set in the current rolling prediction time domain in order to enable the vehicle to meet the green wave traffic constraint (i.e., pass through the intersection ahead without stopping).

[0018] Next, a multi-target phase scan is performed. Within each prediction step, the algorithm iteratively searches for the positions of all traffic lights in the current road segment. If the predicted location does not cross the endpoint, then the nearest traffic light is locked as the current simulation target. Then, the dynamic vehicle speed is calculated, indexed for each discrete time step in the prediction sequence. According to the intersection Real-time transmitted signal phase timing information, also known as SPaT information, is used to calculate a feasible speed range sequence that ensures a single vehicle can pass without stopping. .

[0019] The first step is to calculate the maximum speed required to ensure the vehicle arrives during the green light period, based on the remaining time of the green light and the starting time lost due to the backlog of vehicles ahead. This determines the feasible speed limit, also known as the upper boundary. ), ,in, The maximum speed limit allowed for vehicles on the road; This represents the remaining distance between the current vehicle and the stop line at the traffic light ahead (or the rear of the queue of vehicles). The moment when the green light window opens ahead; This represents the time offset of the prediction point relative to the current time.

[0020] The second step is to consider the end time of the green light and calculate the slowest speed required to ensure that a bicycle can cross the line at the end of the current wave. This is the feasible lower speed limit, also known as the lower boundary. ), When it is predicted that the light ahead is red, virtual guidance based on smooth deceleration is introduced. ,in, The time when the green light window closes; The preset maximum comfort deceleration constant (a positive value is used for calculation); This is the distance from the stop line; To allow for a safe distance when stopping at a red light.

[0021] To eliminate numerical instability caused by spatiotemporal jumps in vehicle speed boundaries and to provide energy-saving guidance for energy management strategies, a guided speed feature vector was constructed. The arithmetic mean of the feasible upper and lower speed limits at each discrete time point was calculated to obtain the guided speed reference value for that discrete time point. In one embodiment, for the predicted 200 discrete speed envelopes, each discrete time point corresponds to one discrete speed envelope. The arithmetic mean trajectory of the feasible upper and lower speed limits corresponding to each discrete speed envelope was calculated to obtain the guided speed reference value for the i-th discrete time point. If the intersection cannot be traversed within the current time domain, the preset minimum cruising speed will be used as the guidance benchmark.

[0022] To meet the high real-time performance and low redundancy requirements of reinforcement learning models for feature observation, instead of directly using the complete 200-step trajectory, non-uniform dimensionality reduction sampling is performed. Multiple guidance velocity reference values ​​are arranged in chronological order, and the guidance velocity feature vector is determined through downsampling. Specifically, this is achieved by extracting future key nodes (such as...) The vehicle speed guidance, i.e., the guidance speed reference value, is used to construct a five-dimensional feature vector. This five-dimensional vector sequence constitutes a predictive observation set, capturing the trend of vehicle speed fluctuations caused by future intersection phase transitions. By loading this trend into the agent's observation space, the model can predict whether the vehicle will be in an acceleration or deceleration phase in 5-20 seconds, thus guiding the hybrid system to allocate power torque in advance and improving the long-term fuel economy of the strategy.

[0023] Conventional methods often directly output specific target vehicle speeds or desired accelerations. These output values ​​are prone to abrupt changes near traffic light phase switching thresholds, leading to uneven control commands and difficulty in conveying long-term speed trends to the energy management layer. The technical solution in this specification constructs a feasible domain sequence containing upper and lower feasible speed limits, and further extracts the arithmetic mean of these limits as a guide speed reference value. This effectively filters out boundary spikes caused by phase transitions, resulting in a temporally smooth guide speed feature vector. Downsampling compresses the high-density time series into a low-dimensional feature vector, preserving future speed trends (e.g., whether acceleration is imminent or deceleration is required) while significantly reducing the dimensionality of the observation space, avoiding the curse of dimensionality. Furthermore, when a vehicle cannot smoothly pass through an intersection, a safe guide speed is automatically calculated based on the maximum comfortable deceleration, and both the upper and lower limits are set to this value. This ensures that even in scenarios with red lights or insufficient passage windows, the guide speed remains physically feasible and comfortable, without providing infeasible speed suggestions exceeding the vehicle's braking capacity.

[0024] Step S102: Extract the information reception delay from the traffic light phase timing information, and determine the multimodal observation state vector based on the current motion state data, the guidance velocity feature vector, and the information reception delay.

[0025] Based on the current motion state data, the guidance speed feature vector, and the information reception delay, a multimodal observation state vector is determined. Specifically, this includes: extracting the current battery state of charge, current vehicle speed, and current driving torque demand from the current motion state data; concatenating the current battery state of charge, current vehicle speed, current driving torque demand, guidance speed feature vector, and information reception delay into an original observation vector according to a preset feature order; and performing linear mapping normalization processing based on a preset feature value interval on each feature dimension of the original observation vector to map each feature value to a unified standardized interval, thereby obtaining the multimodal observation state vector.

[0026] In one embodiment of this specification, information reception delay is extracted from traffic light phase timing information, heterogeneous feature fusion is performed between the vehicle's internal dynamic state and external traffic environment information, and a multimodal observation state vector is determined based on the current motion state data, guidance speed feature vector, and information reception delay to form the decision-making basis of the deep reinforcement learning agent. The specific implementation process is as follows: First, an observation vector containing both internal and external multimodal features is constructed. , Extract the current battery state of charge, current vehicle speed, and current drive torque demand from the current motion data. The vehicle-side observation includes the current battery state of charge. Current driving torque demand And the current actual vehicle speed, i.e., the current vehicle speed Cloud-based observation by sum vector The structure involves extracting information reception delay from the traffic light phase timing information. The time difference between the currently received SPaT information and the current moment is recorded in real time during the information reception delay. If The large value indicates that the data is outdated. The vector refers to the five green wave suggested vehicle speed anchor points obtained in step S101, which reflect the temporal distribution characteristics within the future 0s to 20s period, i.e., the guiding speed feature vector.

[0027] Next, heterogeneous feature fusion and normalization are performed. During feature vector fusion, the low-frequency, slowly varying SOC signal, the high-frequency, fluctuating torque demand, and the guiding speed feature vector containing future trends are logically concatenated. This transforms the originally scattered and asynchronous observations into a unified-dimensional input vector. Due to the difference in the magnitude of the data from different modalities, a state normalization layer is introduced. By performing interval mapping on the input data, the training convergence difficulties caused by the mismatch in magnitude are effectively alleviated, and the sensitivity of the reinforcement learning network to small boundary fluctuations is improved.

[0028] In one embodiment, a preset feature value interval corresponding to each feature dimension is pre-stored, i.e., the minimum and maximum values ​​of the feature within its normal operating range. During normalization, the actual value of the feature is subtracted from the minimum value of the interval, and then divided by the interval length to obtain the mapped ratio. This ratio is then multiplied by the length of the target standardized interval and the offset of the standardized interval is added, thereby linearly transforming the original value to a unified standardized interval (e.g., between -1 and +1). After all feature dimensions have undergone the above mapping independently, the resulting new vector is the multimodal observation state vector. Each element in this multimodal observation state vector has a uniform scale, which can be directly used as the input to neurons in the deep reinforcement learning network, avoiding the adverse effects of dimensional differences on network parameter updates.

[0029] The above technical solution integrates the vehicle's own power domain state (battery state of charge, vehicle speed, torque demand) with external environmental domain features (guided speed trend, communication latency) into a single observation vector. Linear mapping normalization unifies all features to the same numerical range, eliminating dimensional interference and enabling the reinforcement learning agent to learn the weight of each feature's influence on decision-making equally. This accelerates network convergence and improves the generalization performance of the policy. Furthermore, by explicitly inputting information reception latency as an observation feature, the agent can perceive the quality of current vehicle-to-infrastructure communication. When communication latency is high, the agent can automatically reduce its trust in the guided speed feature vector and adjust its decision-making behavior based on subsequent confidence coefficients, thereby enhancing the system's robustness in non-ideal communication environments. Simultaneously, concatenating the original observation vector in a fixed order and uniformly normalizing it forms a standardized preprocessing pipeline, facilitating efficient and deterministic computation on an in-vehicle embedded platform and meeting the real-time control requirements for processing latency.

[0030] Step S103: Input the multimodal observation state vector into the probabilistic action network in the pre-trained soft actor-critic network, and output the mean parameter and standard deviation parameter to describe the probability distribution of the equivalent factor. Calculate the reliability coefficient based on the standard deviation parameter, and use the reliability coefficient to weight and fuse the preset dual-strategy equivalent factors to generate the target equivalent factor.

[0031] Before inputting the multimodal observation state vector into the probabilistic action network in the pre-trained soft actor-critic network in step S103, the method further includes: acquiring raw driving data containing multiple historical driving trajectories, extracting speed and acceleration states from the raw driving data, constructing a two-dimensional discrete state space, and calculating a first-order Markov state transition probability matrix; generating a basic speed trajectory based on the first-order Markov state transition probability matrix using the Monte Carlo method, and determining whether the basic speed trajectory meets the constraints according to the average speed access interval of the green wave traffic segment; if not, performing a local mutation operation on the trajectory segment that does not meet the constraints to generate multiple candidate mutation segments; calculating the comprehensive evaluation score of each candidate mutation segment, replacing the original trajectory segment with the candidate mutation segment corresponding to the highest value in the comprehensive evaluation score, and iteratively generating a green wave traffic condition library; and training the soft actor-critic network using the green wave traffic condition library.

[0032] In one embodiment of this specification, to address the lack of specific green wave traffic condition data in reinforcement learning model training, a high-fidelity training condition library is constructed by combining the statistical randomness of human driving behavior with green wave traffic constraints. Then, the soft actor-commentator network is trained using this green wave traffic condition library. The implementation process for constructing the green wave traffic condition library is as follows: First, raw driving data containing multiple historical driving trajectories is acquired. This raw driving data is based on real-vehicle driving trajectory records, and the collection scenarios cover urban main roads and typical green wave traffic light networks, possessing high behavioral representativeness. Noise reduction processing is applied to the raw driving trajectories, and speed and acceleration states are extracted from the raw driving data. Single-vehicle speed *v* and acceleration *a* are selected as key state variables representing driving habits, constructing a two-dimensional discrete state space. The first-order Markov state transition probability matrix (STPM) is constructed by statistically analyzing the transition frequencies within a given state space. Elements in the matrix Indicates the bicycle's current state Transition to the next state The probability of is calculated using the following formula: , in, Represents the state observed in the original dataset. to state The transition frequency of this matrix The random rhythms and driving characteristics of human drivers are digitally encoded.

[0033] For green wave scenarios, vehicle trajectories not only need to be realistic, but also must comply with green wave traffic constraints, such as... Figure 3 As shown, Figure 3 This document provides a schematic diagram of traffic light phase constraints and vehicle spatiotemporal trajectories as an embodiment of the present specification. To generate high-precision operating conditions that combine human driving characteristics with green wave traffic constraints, this document proposes a closed-loop iterative strategy involving Markov generation, local mutation, and weight optimization.

[0034] First, using the Monte Carlo method, a basic speed trajectory is generated based on the first-order Markov state transition probability matrix. The constraint on whether the basic speed trajectory meets the requirements is then determined based on the average speed access interval of the green wave traffic segment. Specifically, the constructed first-order Markov state transition probability matrix is ​​used... The Monte Carlo method was used to sample and generate the baseline velocity trajectory. Green wave traffic sections. Traffic constraints are determined by the traffic light phase difference and bandwidth. The average speed allowable zone for this road segment is defined as follows: ,in, This is the minimum average speed limit for this road section; vehicles below this value cannot reach the intersection before the green light zone. This represents the average speed limit for this road segment; vehicles exceeding this value will arrive at the intersection before the green light period begins. The generated base speed trajectory is located on the road segment. Average speed within Must meet To ensure that vehicles enter the intersection during the green light phase, allowing for instantaneous speeds. It exceeds the mean range.

[0035] If the constraints are not met, a local mutation operation is performed on the trajectory segments that do not meet the constraints, generating multiple candidate mutated segments. That is, when the synthesized trajectory of a road segment exceeds the green wave traffic constraints, the algorithm automatically triggers mutation, extracts the affected trajectory segments, and regenerates multiple sets of candidate mutated segments that satisfy the characteristics of the first-order Markov state transition probability matrix and whose trends are subject to boundary constraints. A comprehensive evaluation score is calculated for each candidate mutated segment, and the original trajectory segment is replaced with the candidate mutated segment corresponding to the highest comprehensive evaluation score, iteratively generating a green wave traffic condition library. Here, the comprehensive evaluation score is defined as follows: To select the optimal mutation trajectory. ,in, , These are preset weighting coefficients that reflect the importance of central tendency and likelihood. The average vehicle speed of the generated candidate variant trajectory segments; The recommended center trajectory for balanced vehicle speed in the current green wave window. The Markov likelihood is used to evaluate the degree of fit between the generated sequence and the original first-order Markov state transition probability matrix, ensuring the authenticity of the driving style. Low-scoring non-steady-state trajectories are dynamically eliminated, and the optimal candidate path is stitched together and reconstructed. The high-fidelity output driving conditions ensure that a single vehicle has a high adaptability to green wave scenarios with different road sections and signal timings during large-scale generation, providing driving condition data support for the offline training of the energy management agent in the following section.

[0036] The randomness of human driving behavior is modeled using a first-order Markov state transition probability matrix. Then, green wave traffic constraints are embedded into the trajectory generation process using a local mutation operator, so that the generated driving conditions retain the speed of the real driver. Acceleration statistics are used to enforce compliance with green wave window requirements. By employing a comprehensive evaluation score to select the best candidate variant segments, the average speed of the generated trajectory is ensured to approximate the suggested center speed of the green wave, and the statistical characteristics of state transitions are consistent with real driving, avoiding trajectory distortion caused by abrupt truncation or forced assignment. The iterative optimization closed-loop mechanism ensures that every trajectory in the large-scale driving condition library is statistically similar to real driving and strictly compliant with green wave constraints. Reinforcement learning agents trained using this dedicated driving condition library exhibit stronger adaptability and more stable energy management performance when facing varying intersection spacing, different traffic light timing schemes, and random traffic interference, significantly narrowing the performance gap between simulation training and real-world applications.

[0037] Next, the soft actor-critic network is trained using a green wave traffic condition database. The reward function during training includes a fuel consumption quality weighting term, a battery state-of-charge deviation weighting term, and an expert knowledge penalty term. The expert knowledge penalty term is the difference between the fuel consumption rate at the current engine operating point and the optimal fuel consumption rate at the corresponding engine speed. The expert knowledge penalty term is multiplied by a preset adjustment coefficient and then added to the fuel consumption quality weighting term and the battery state-of-charge deviation weighting term to determine the total reward signal. The soft actor-critic network updates its parameters based on the total reward signal by maximizing the expected cumulative reward, enabling the probabilistic action network to learn and output an equivalent factor probability distribution that maximizes the total reward.

[0038] In one embodiment of this specification, an improved soft actor-critic algorithm is employed to achieve adaptive optimization and safety control of the equivalence factor by real-time perception of environmental uncertainty. Unlike traditional reinforcement learning that outputs fixed scalar values, this embodiment uses a probabilistic neural network architecture. The policy network (Actor) no longer directly outputs a single, definite value for the equivalence factor, but instead outputs a pair of parameters describing the distribution of the equivalence factor, including the mean. and standard deviation Mean The standard deviation represents the algorithm's optimal decision preference under the current operating conditions. This represents the algorithm's uncertainty regarding the current decision. The standard deviation branch is handled using the Softplus activation function at the network's end to ensure... When V2I communication is delayed or traffic light phase timing information (SPaT information) is lost, the network outputs a larger... This is used to characterize the high uncertainty of the environment.

[0039] Introducing an expert knowledge penalty term into the reward function guides the agent's learning. ,in, It is the fuel consumption weighting coefficient. For fuel consumption quality, It is the SOC deviation weighting coefficient. It is the adjustment coefficient for the expert knowledge penalty item. For expert knowledge penalty items, ,in The engine output torque at that point. This represents the engine output speed at that point. This is the Brake Specific Fuel Consumption (BSFC) value at that point. This is the minimum BSFC value among different torques at that speed. This penalty term allows the agent to quickly converge to an efficient region that conforms to engineering logic during offline training.

[0040] The Soft Actor-Critic Network employs a maximum entropy reinforcement learning framework, aiming not only to maximize the expected cumulative reward but also to maximize the policy entropy. The Soft Actor-Critic Network consists of a critic network (for evaluating state-action value) and a probabilistic action network (for outputting the action distribution). The observed state, action (i.e., the actual sampled equivalent factor), total reward signal, and next observed state at each time step are stored in an experience replay buffer. At each network parameter update, a batch of experience data is randomly sampled from the buffer, the temporal difference error is calculated to update the critic network, and the probabilistic action network is updated by minimizing the negative expected reward plus an entropy regularization term. After extensive iterative training, the probabilistic action network learns the equivalent factor probability distribution that maximizes the expected total reward signal at a given observed state. This distribution minimizes fuel consumption and guides the engine to operate in its high-efficiency range while ensuring stable battery state of charge.

[0041] By introducing an expert knowledge penalty term, the optimal efficiency curve from the engine's universal characteristic graph is explicitly encoded into the reward function. This allows the agent to naturally tend to choose equivalent factors that bring the engine closer to its high-efficiency zone during the learning process, thereby achieving better overall powertrain efficiency at the same fuel consumption level. Since the expert knowledge penalty term is dynamically calculated (varying with engine speed), the agent can learn to distinguish the differences in the high-efficiency zone at different engine speeds, rather than rigidly pursuing a fixed torque. Training using a dedicated green wave traffic condition library ensures that the agent is exposed to rich combinations of traffic light phase timings and speed trajectory changes during the training phase, significantly shortening the transition from simulation to real-world driving. This enables the probabilistic action network to output effective strategies that satisfy fuel economy, state-of-charge balance, and engine operating point efficiency, providing a reliable and high-performance decision-making basis for online execution.

[0042] In one embodiment of this specification, a multimodal observation state vector is input into a probabilistic action network in a pre-trained soft actor-critic network, and the output is a mean parameter and a standard deviation parameter describing the probability distribution of the equivalent factor. The reliability coefficient is calculated based on the standard deviation parameter, and the preset dual-strategy equivalent factors are weighted and fused using the reliability coefficient to generate the target equivalent factor.

[0043] The reliability coefficient is calculated based on the standard deviation parameter, specifically including: obtaining the standard deviation parameter output by the probabilistic action network and reading a preset uncertainty threshold; determining the relationship between the standard deviation parameter and the uncertainty threshold; if the standard deviation parameter is less than the uncertainty threshold, setting the reliability coefficient to a first preset constant; if the standard deviation parameter is not less than the uncertainty threshold, calculating an exponential function with a base of the natural constant and an exponent of the negative of the difference between the standard deviation parameter and the uncertainty threshold, to determine the reliability coefficient.

[0044] Specifically, to ensure that the soft actor-critic (SAC) algorithm can completely dominate under normal operating conditions, an uncertainty threshold is introduced. Reliability function The reliability function can also be called the reliability coefficient. ：

[0045] in, This is a preset confidence threshold. The variance of the agent's output... When the situation is within the threshold (i.e., less than the uncertainty threshold and greater than 0), the current environment is considered to be in the high confidence interval. At this point, the confidence function... Forced to be set to 1, completely eliminating the bias interference of the baseline strategy on the optimization results; when When the threshold is exceeded, the confidence function As uncertainty increases, the decay rate decreases exponentially, triggering the security protection logic.

[0046] The target equivalent factor is generated by weighting and fusing the preset dual-strategy equivalent factors using the reliability coefficient. Specifically, this includes: obtaining a first strategy equivalent factor and a second strategy equivalent factor from the dual-strategy equivalent factors; wherein the first strategy equivalent factor is determined by the mean parameter output by the probabilistic action network, and the second strategy equivalent factor is calculated by the adaptive equivalent fuel consumption minimization strategy based on the current battery state of charge; determining a first weighted component by multiplying the reliability coefficient by the first strategy equivalent factor; calculating the difference between the first and second strategies and multiplying the difference by the second strategy equivalent factor to determine a second weighted component; and determining the target equivalent factor based on the first and second weighted components.

[0047] Specifically, the system automatically and smoothly switches between strategies based on the reliability coefficient corresponding to the environmental reliability, ultimately outputting an adaptive equivalent factor. That is, the target equivalence factor, the formula is: ,in, It is an equivalent factor value calculated based on the SAC algorithm. It is a conservative reference value calculated based on the classic SOC balance (AECMS). When the signal environment is excellent ( When communication is abnormal or the operating conditions are extremely complex, the fuel-saving strategy output by SAC is executed; when communication is abnormal or the operating conditions are extremely complex, When the system is in operation, it automatically reverts to conservative control based on AECMS to ensure SOC balance and driving safety.

[0048] Normal / high confidence scenarios ( The system fully implements the SAC policy. ),at this time This ensures that the system can maximize fuel-saving potential when vehicle-road cooperative information is accurate. (Anomaly / High Uncertainty Scenarios) When V2I communication failure or other factors cause the agent to malfunction, Rapid reduction and automatic degradation to conservative control based on A-ECMS ensures battery SOC balance and driving safety under extremely complex operating conditions.

[0049] The system quantifies environmental uncertainty in real time using the standard deviation parameter output by the soft actor-critic network and dynamically generates a reliability coefficient accordingly. This enables the system to autonomously identify when to trust the reinforcement learning strategy and when to switch to a conservative benchmark, avoiding the lack of awareness of communication quality in traditional methods. An exponential decay method is used to calculate the reliability coefficient, achieving smooth and rapid weight reduction in high uncertainty intervals and avoiding control abrupt changes caused by hard switching. Weighted fusion ensures the continuity of the transition between the two strategies, making the change of the equivalent factor smoother and beneficial to the stable operation of the hybrid power system. The system addresses the control robustness problem when vehicle-road cooperative information is not ideally available in green wave traffic scenarios, significantly improving the reliability of energy management strategies in real urban traffic environments while retaining the fuel economy advantages of deep reinforcement learning strategies under good communication conditions.

[0050] Step S104: Based on the current vehicle speed, current driving torque demand, and target equivalent factor of the current motion state, construct an evaluation index function for instantaneous equivalent energy consumption, and solve for the engine target torque command and motor target torque command that minimize the evaluation index function within the preset powertrain physical constraint feasible set, so as to realize hybrid vehicle energy management.

[0051] In one embodiment of this specification, the equivalent factor s(t) output by the soft actor-critic network is used to perform millisecond-level dynamic allocation while satisfying physical constraints. This is based on an adaptive equivalent factor. Real-time construction of evaluation index for minimizing instantaneous equivalent energy consumption, Hamiltonian function The Hamiltonian function is the evaluation index function: , in, and These are engine torque and speed, respectively; and These are the motor torque and speed, respectively. This refers to the output power at the battery end. This is the lower calorific value coefficient of fuel. This function transforms the global fuel optimization problem into an instantaneous optimal trade-off problem. Equivalent factor. As a dynamic weight for oil-electricity balance, it reflects the immediate value of electricity under the current green wave environment and SOC trend. At this point, the adaptive equivalent factor determined by SAC... It has incorporated the green wave guidance trend for the next 20 seconds, ensuring that the evaluation criteria for the Hamiltonian function have dynamic predictive capabilities.

[0052] To obtain real-time power load demand (torque demand) After that, the underlying controller, within the feasible control set that satisfies the hardware safety limits... Perform optimal control vector within Global search, To optimize the engine target torque command, This is the target torque command for the motor.

[0053] Within a pre-defined feasible set of powertrain physical constraints, the engine target torque command and motor target torque command that minimize the evaluation index function are solved. During this process, hard constraint limiting is applied, and strict physical boundary protection is implemented at the execution level. Simultaneously, the battery side is constrained by the battery peak charge / discharge power (SOP) and the current SOC current limit, while the power side is constrained by the engine steady-state speed range, dynamic torque response limitations, and the motor field weakening limitation. This is achieved through a convex / non-convex optimization interval defined by the above constraints. Internal search Hamiltonian function The minimum point is obtained, which improves the risk of instability that may be caused by random exploration by the agent from the algorithm level.

[0054] Specifically, the feasible set of powertrain physical constraints consists of hardware boundary conditions pre-stored in the controller. These include: the minimum and maximum speed range for steady-state engine operation; limits on the rate of increase and decrease of engine dynamic torque; maximum torque limits in motor drive and generator modes (affected by motor cooling conditions and current speed, respectively); peak charging power limits and peak discharging power limits for the battery (calculated in real-time by the battery management system based on the current battery state of charge and temperature); and the equation constraint that the total wheel-end torque determined by the current drive torque demand must equal the sum of the engine output torque and the motor output torque (after reduction by the reducer and main reduction ratio). An enumeration scan or numerical optimization algorithm is used to discretize and sample the engine torque within the feasible set. The possible engine output torque is divided into multiple candidate values ​​from the minimum to the maximum allowable value with a fixed step size. For each candidate engine torque, the corresponding motor torque is calculated according to the torque balance equation. Then, it is checked whether the motor torque falls within the motor torque limit range. Simultaneously, based on the current battery state of charge and temperature, it is calculated whether the motor power exceeds the battery's charging and discharging power limits. If all constraints are met, the candidate combination is substituted into the evaluation index function to calculate the instantaneous equivalent energy consumption value; if not, it is skipped. After traversing all feasible candidate combinations, the engine torque that minimizes the evaluation index function is selected as the engine target torque command, and the corresponding motor torque is selected as the motor target torque command. Finally, the engine target torque command is sent to the engine controller and the motor target torque command is sent to the motor controller via the controller area network bus or vehicle Ethernet. The underlying actuators respond to the commands, completing the energy distribution and vehicle drive for this control cycle, thus realizing energy management of hybrid vehicles in green wave traffic scenarios.

[0055] Employing a layered architecture, the reinforcement learning agent does not directly control torque. Instead, it outputs an equivalent factor as a scalar. A lower-level optimizer, based on a strategy to minimize equivalent fuel consumption, then solves for the optimal torque combination within a strictly physically constrained feasible set. This separates the flexibility of the learning strategy from the safety of deterministic optimization, fundamentally preventing the output of illegal control commands. Because the Hamiltonian function introduces a target equivalent factor, which carries the global decision-making intent after compromising between upper-level guidance information (green wave speed trend) and communication uncertainties, the lower-level optimization is not a blind, instantaneous optimization, but rather an instantaneous trade-off based on an understanding of future energy consumption trends, achieving a unity of foresight and real-time performance. Simultaneously, by systematically constraining the power and torque boundaries of the engine, motor, and battery, each candidate command undergoes hardware limit verification, ensuring the physical realizability of the commands and avoiding control deviations caused by incomplete execution. The final output of the engine and motor target torque commands is standardized in format and can be directly recognized and executed by the lower-level controller, forming a fully automatic closed loop from traffic information input to power output. It leverages the advantages of deep reinforcement learning in handling complex, high-dimensional dynamic environments while retaining the reliability of model-based optimization methods in terms of physical constraint handling and real-time performance, making it suitable for hybrid vehicles with stringent safety and economic requirements.

[0056] The at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: By extracting the guiding speed feature vector through rolling time-domain extrapolation, the spatiotemporal constraints of the intersection traffic lights are transformed into a quantitative representation of future speed trends, enabling the energy management strategy to be predictive and actively adjust the battery state of charge before the green light window arrives, reducing the intervention of the engine in the inefficient operating range, thereby significantly improving fuel economy. Utilizing the probability distribution of the equivalent factor output by the soft actor-critic network, environmental uncertainty is quantified in real time using the standard deviation parameter, and a confidence coefficient is calculated based on this. This achieves adaptive fusion decision-making that fully exploits the agent's optimization potential under high confidence and automatically degenerates into a conservative state of charge maintenance strategy under high uncertainty, greatly enhancing robustness and safety in complex real-world environments. Adopting a hierarchical architecture, the equivalent factor output by the agent is used as the weight parameter of the lower-level equivalent fuel consumption minimization strategy. The optimal torque combination is solved within the feasible set of strict powertrain physical constraints, preserving the global optimization capability of the data-driven strategy while ensuring the physical realizability and engineering safety of each set of instructions. Hybrid vehicles that continuously reduce energy consumption while ensuring driving safety and component reliability are suitable for green wave traffic scenarios.

[0057] Figure 4This specification provides a schematic diagram of the system architecture of a hybrid vehicle energy management system based on green wave traffic, illustrating an embodiment of the system. It constructs a hierarchical management architecture that integrates vehicle-road cooperative guidance information with artificial intelligence decision-making, achieving a deep integration of traffic guidance information and underlying power torque distribution. Figure 4 The diagram illustrates a three-layer control architecture encompassing perception, decision-making, and execution. Specifically, the upper layer utilizes V2I technology to transform the spatiotemporal constraints of traffic lights into vehicle speed guidance; the middle layer, the SAC (System-Agent Control) agent, adjusts the equivalent factor in real time based on multimodal observation features; and the lower layer, the deterministic control layer ECMS, outputs torque commands under feasible set constraints. This layered architecture enables the powertrain to predictively perceive the green wave window ahead, effectively ensuring that the vehicle operates within its optimal efficiency range in complex dynamic environments.

[0058] Figure 5 This is a schematic diagram of the overall process of a layered strategy provided in an embodiment of this specification, such as... Figure 5 As shown, a green wave working condition construction system integrating local mutation algorithm is first established, which provides a training working condition library with high generalization and high fidelity for the model.

[0059] Specifically, based on the actual collected driving data, the joint state distribution of speed and acceleration is extracted, and the statistical characteristics of single-vehicle driving behavior are modeled using a first-order Markov process. Building upon this, and combining the green wave traffic constraint information provided by V2I, the generated trajectory is reconstructed using a local mutation algorithm. Addressing the lack of green wave traffic features in standard driving conditions (such as WLTC), a method for constructing driving conditions that integrates random driving behavior and green wave traffic constraints is proposed. By capturing the statistical regularities of basic driver driving behavior through Markov stochastic processes and using a local mutation algorithm to correct the original trajectory based on green wave traffic constraints, a large-scale driving condition library covering diverse intersections and dynamic traffic features is constructed. This effectively reduces the policy distribution deviation between simulation training and real-world vehicle operation, ensuring the robustness and decision stability of the energy management system in the face of random traffic disturbances; simultaneously, it ensures that the vehicle speed sequence retains the randomness of real driving behavior while conforming to the spatiotemporal window constraints of green wave traffic at intersections, providing a rich training condition library for the agent.

[0060] To address the shortcomings of existing deep reinforcement learning observation spaces in capturing traffic guidance information and lacking predictive capabilities, a rolling time-domain algorithm was used to extract guidance sequences characterizing the trend of green wave traffic speed. Utilizing vehicle-to-infrastructure (V2I) interaction technology, the signal phase and timing (SPaT) information at the upcoming intersection is analyzed in real time. Based on the vehicle's current motion state, a rolling time-domain extrapolation algorithm provides the vehicle with a feasible reference speed domain. (in Lower bound of the feasible region velocity (This serves as the upper bound of the feasible region speed) to ensure smooth, uninterrupted passage of vehicles within the green wave window. To further mitigate the instantaneous abrupt changes in decision-making instructions under complex traffic conditions, reference speed feature vectors at multiple time scales are extracted. This feature is taken as the instantaneous arithmetic mean of the upper and lower bounds of the feasible region, covering the evolution trend within the next 20 seconds, and is used as the input of the guiding observation term into the reinforcement learning observation space. The dynamic vehicle speed range under the traffic light constraint for the next 20 seconds is digitally represented, enabling the SAC agent to have a long-term observation perspective. It can pre-charge or pre-discharge the battery state of charge (SOC) in advance based on the spatiotemporal window information of the traffic light ahead, mitigating the decrease in energy utilization efficiency caused by sudden speed changes.

[0061] Then, a multimodal enhanced observation space was constructed. This includes the battery state of charge (SOC). Instantaneous vehicle speed and driving torque requirements () The real-time status of the vehicle reflects its dynamic characteristics, and also includes the dynamic time delay of the response signal status. and the velocity-guided feature vector for the next 20 seconds extracted from the aforementioned steps, serving as the predicted state. ( Depend on (Obtained through dimensionality reduction). Reinforcement learning agents can perceive speed fluctuations at intersections in advance, providing a global decision-making basis for adjusting battery SOC balance and engine efficiency.

[0062] Simultaneously, a hierarchical SAC and ECMS policy architecture integrating uncertainty perception and expert knowledge heuristics is proposed, improving the robustness and stability of the policy under complex and variable operating conditions. A closed-loop decision-making system from intelligent control to deterministic execution is constructed, employing an improved soft actor-critic (SAC) algorithm, and synchronously outputting the mean describing the distribution of equivalent factors through a probabilistic actor network. with standard deviation ,use Real-time quantification of environmental uncertainties caused by V2I communication latency, packet loss, etc., and introduction of expert knowledge penalty term into the reward function. Using the engine's optimal efficiency line as a benchmark, the agent is guided to quickly converge to the engineering efficiency range during the training phase; further, a dynamic reliability evaluation function with a confidence threshold is introduced, where the standard deviation... At the preset threshold Within a certain range, the reliability coefficient The value is always set to 1 to ensure that the SAC strategy executes independently in an undisturbed state; if environmental uncertainty exceeds the threshold, then... Follow Increases and then decays exponentially ( ), through weighted fusion logic This enables a smooth switch between SAC deep optimization and AECMS conservative benchmark, ensuring automatic degradation when information is unreliable to guarantee SOC balance and driving safety. The lower control layer receives the synthesized adaptive equivalent factor. Real-time construction of instantaneous Hamiltonian functions The control feasible set is based on the universal characteristics of the engine, the MAP diagram of the motor, and the physical boundary of the battery. The optimal engine and motor torque command vector is obtained by minimizing the input. By utilizing physical deterministic laws to constrain and limit the decision-making of intelligent agents, the powertrain is ensured to always operate in the optimal efficiency range under complex green wave environments, balancing the exploration flexibility of the algorithm with the robustness of practical engineering.

[0063] Figure 6 This diagram illustrates the SOC maintenance effect of an embodiment provided in this specification. It compares the SOC maintenance and fuel economy effects of the V2I-SAC-ECMS method (integrating vehicle-road cooperative information and artificial intelligence decision-making) and the AECMS method under the same green wave operating conditions. The AECMS solution has undergone multiple rounds of precise calibration and parameter coverage testing, achieving a superior energy efficiency level under feedback control, and thus possesses high practical value and comparative significance. Figure 6 It can be seen that, under the same operating conditions, the SOC fluctuation of the V2I-SAC-ECMS method provided in the embodiments of this specification is smaller than that of the AECMS method. In other words, the embodiments of this specification have better SOC maintenance stability. Figure 7 This is a schematic diagram of an engine operating point diagram for fuel economy provided in an embodiment of this specification, as shown below. Figure 7 As shown, simulation results indicate that the AECMS strategy has a final soc of 48.70% and a fuel consumption of 0.367L; the V2I-SAC-ECMS strategy has a final soc of 45.69% and a fuel consumption of 0.337L. After converting the energy consumption, the V2I-SAC-ECMS strategy reduces fuel consumption by approximately 6.7% compared to AECMS, demonstrating better fuel economy.

[0064] This specification also provides an embodiment of a hybrid vehicle energy management device based on green wave traffic, such as... Figure 8 As shown, the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described method.

[0065] This specification also provides a non-volatile computer storage medium storing computer-executable instructions configured to perform the above-described method.

[0066] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0067] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A hybrid vehicle energy management method based on green wave passing, characterized by, The method includes: The system acquires the current motion state data of the hybrid vehicle and obtains the traffic light phase timing information of the current intersection through the vehicle-road communication module. Based on the traffic light phase timing information and the current vehicle speed in the current motion state data, it determines a guiding speed feature vector to characterize the future speed trend. Information reception delay is extracted from the traffic light phase timing information, and a multimodal observation state vector is determined based on the current motion state data, the guidance speed feature vector, and the information reception delay. The multimodal observation state vector is input into the probabilistic action network in the pre-trained soft actor-critic network, and the mean parameter and standard deviation parameter are output to describe the probability distribution of the equivalent factor. The reliability coefficient is calculated based on the standard deviation parameter, and the preset dual-strategy equivalent factors are weighted and fused using the reliability coefficient to generate the target equivalent factor. Based on the current vehicle speed, current driving torque demand, and target equivalent factor of the current motion state, an evaluation index function for instantaneous equivalent energy consumption is constructed. The engine target torque command and motor target torque command that minimize the evaluation index function are solved within the preset powertrain physical constraint feasible set to realize hybrid vehicle energy management. Before inputting the multimodal observation state vector into the probabilistic action network in the pre-trained soft actor-critic network, the method further includes: Obtain raw driving data containing multiple historical driving trajectories, extract speed and acceleration states from the raw driving data, construct a two-dimensional discrete state space, and calculate the first-order Markov state transition probability matrix. Using the Monte Carlo method, a basic speed trajectory is generated based on the first-order Markov state transition probability matrix. The basic speed trajectory is then judged to meet the constraints based on the average speed access interval of the green wave traffic segment. If the constraints are not met, a local mutation operation is performed on the trajectory segments that do not meet the constraints, generating multiple candidate mutated segments; Calculate the comprehensive evaluation score for each candidate variant segment, replace the original trajectory segment with the candidate variant segment corresponding to the highest value among the comprehensive evaluation scores, and iteratively generate the green wave traffic condition library; The soft actor-critic network is trained using the green wave traffic condition database.

2. The hybrid vehicle energy management method based on green wave traffic according to claim 1, characterized in that, Based on the traffic light phase timing information and the current vehicle speed in the current motion state data, a guiding speed feature vector is determined to characterize future speed trends, specifically including: Based on the traffic light phase timing information and the current vehicle speed, the feasible upper and lower speed limits of the hybrid vehicle at each discrete time point within a preset future time period are predicted, and the sequence of feasible speed domains is determined. The arithmetic mean of the feasible upper and lower speed limits at each discrete time point is used to obtain the reference value of the guiding speed at each discrete time point. After arranging the multiple guidance speed reference values ​​in chronological order, the guidance speed feature vector is determined by downsampling.

3. The hybrid vehicle energy management method based on green wave traffic according to claim 2, characterized in that, Based on the traffic light phase timing information and the current vehicle speed, the feasible upper and lower speed limits of the hybrid vehicle at each discrete time point within a preset future time period are predicted, and the feasible speed domain sequence is determined, specifically including: Starting from the current sampling time, iterate forward at discrete time points with a preset time step until the preset future duration is reached; At each discrete time point, it is determined whether the virtual position of the vehicle has crossed the stop line of the traffic light ahead based on the traffic light phase timing information; If the crossing is not completed, the upper and lower limits of the feasible speed are calculated based on the green light start and end times in the current traffic light phase timing information, combined with the current position of the hybrid vehicle and the remaining distance between the current position of the hybrid vehicle and the stop line of the traffic light ahead. The upper limit of the feasible speed is the fastest speed at which the hybrid vehicle can reach the stop line after the green light starts, and the lower limit of the feasible speed is the slowest speed at which the hybrid vehicle can reach the stop line before the green light ends. If the stop line has already been crossed, the safe guide speed for the hybrid vehicle to stop before the stop line without crossing the stop line of the traffic light ahead is calculated based on the preset maximum comfort deceleration, and both the upper and lower limits of the feasible speed are set to the safe guide speed.

4. The hybrid vehicle energy management method based on green wave traffic according to claim 1, characterized in that, Based on the current motion state data, the guidance velocity feature vector, and the information reception delay, the multimodal observation state vector is determined, specifically including: Extract the current battery state of charge, current vehicle speed, and current drive torque requirement from the current motion state data; The current battery state of charge, current vehicle speed, current driving torque demand, the guide speed feature vector, and the information reception delay are concatenated into the original observation vector according to a preset feature order. For each feature dimension in the original observation vector, a linear mapping normalization process based on a preset feature value interval is performed to map each feature value to a unified standardized interval, thereby obtaining the multimodal observation state vector.

5. The hybrid vehicle energy management method based on green wave traffic according to claim 1, characterized in that, The reliability coefficient is calculated based on the standard deviation parameter, specifically including: Obtain the standard deviation parameter of the output of the probabilistic action network, and read the preset uncertainty threshold; Determine the relationship between the standard deviation parameter and the uncertainty threshold; If the standard deviation parameter is less than the uncertainty threshold, then the reliability coefficient is set to a first preset constant; If the standard deviation parameter is not less than the uncertainty threshold, then the reliability coefficient is determined by calculating an exponential function with the natural constant as the base and the negative of the difference between the standard deviation parameter and the uncertainty threshold as the exponent.

6. The hybrid vehicle energy management method based on green wave traffic according to claim 1, characterized in that, The target equivalence factor is generated by weighting and fusing the preset dual-strategy equivalence factors using the reliability coefficient, specifically including: Obtain the first strategy equivalence factor and the second strategy equivalence factor from the dual-strategy equivalence factors; Wherein, the first strategy equivalent factor is determined by the mean parameter output by the probabilistic action network, and the second strategy equivalent factor is calculated by the adaptive equivalent fuel consumption minimization strategy based on the current battery state of charge. The first weighted component is determined by multiplying the reliability coefficient by the first strategy equivalence factor. Calculate the difference between the reliability coefficient and the second strategy equivalence factor, and determine the second weighted component by multiplying the difference by the second strategy equivalence factor. The target equivalence factor is determined based on the first weighted component and the second weighted component.

7. The hybrid vehicle energy management method based on green wave traffic according to claim 1, characterized in that, The method further includes: The reward function during the training phase includes a fuel consumption quality weighting term, a battery state of charge deviation weighting term, and an expert knowledge penalty term, wherein the expert knowledge penalty term is the difference between the fuel consumption rate at the current engine operating point and the optimal fuel consumption rate at the corresponding speed. The total reward signal is determined by multiplying the expert knowledge penalty item by a preset adjustment coefficient and then adding it to the fuel consumption quality weighting item and the battery state of charge deviation weighting item. The soft actor-critic network updates its parameters based on the total reward signal by maximizing the expected cumulative reward, enabling the probabilistic action network to learn and output an equivalent factor probability distribution that maximizes the total reward.

8. A hybrid vehicle energy management device based on green wave traffic, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1-7.

9. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to perform the method as described in any one of claims 1-7.

Images