Motor home adaptive energy management system and method based on AI and multi-sensor fusion

Abstract

The invention discloses a limo self-adaptive energy management method based on AI and multi-sensor fusion. The limo self-adaptive energy management method comprises the following steps: acquiring illumination intensity, environment temperature, humidity, wind speed and geographical location information of an external environment of a limo through an environment sensing module; the running speed, the engine start-stop state, the generator output power, the chassis power supply state and the parking state information of the motor home are obtained through the vehicle state monitoring module, the AI reasoning result is directly mapped to the execution unit through the closed-loop control architecture, a physical constraint verification mechanism is built in, the control intelligence is guaranteed, and the control reliability is improved. And the security and robustness of the system are ensured. Finally, the comprehensive efficiency of the motor home energy system is remarkably improved, invalid energy consumption can be reduced by more than 15% under typical working conditions, the cycle life of the lithium ion battery pack is prolonged by more than 20%, and non-perceptual and personalized energy use experience is provided for users.

Description

Technical Field

[0001] This invention relates to the field of RV control technology, specifically to an RV adaptive energy management system and method based on AI and multi-sensor fusion. Background Technology

[0002] With the increasing intelligence and green energy applications in RVs, energy management in mobile environments is facing unprecedented complexity and dynamic challenges. Modern RVs integrate solar power generation, lithium battery storage, mains power access, and various high-power loads (such as air conditioning, refrigerators, and lighting systems). Their energy supply and consumption are affected in real time by multiple factors, including driving status, environmental climate, user behavior, and battery health. Traditional energy management systems are mostly based on fixed rules or threshold logic for control, relying on single electrical parameters (such as voltage and current) for decision-making. They lack the ability to collaboratively perceive and intelligently analyze multi-dimensional physical quantities, making it difficult to achieve efficient, safe, and personalized energy scheduling in highly uncertain mobile environments.

[0003] Among these, adaptive energy management based on multi-sensor fusion and artificial intelligence has become a key direction for improving system performance. This technology aims to build a comprehensive perception of energy supply and demand by integrating heterogeneous data sources such as ambient light intensity, temperature, humidity, vehicle speed, battery SOC / SOH, and user electricity usage habits. It then utilizes lightweight AI models for online learning and prediction to dynamically generate optimal charging / discharging strategies and load allocation schemes. However, existing technologies have not effectively established a closed loop between the perception layer, decision-making layer, and execution layer, resulting in system lag and rigid strategies that cannot meet the adaptive needs of RVs under complex conditions such as long-distance travel, extreme weather, or sudden high loads.

[0004] In existing technologies, some energy management systems, such as CN107274300B, while possessing historical energy consumption clustering and efficiency assessment functions, are designed for fixed sites, relying solely on static rule judgments based on meter data. They lack AI-driven dynamic prediction mechanisms and fail to integrate key dimensions such as environment and user behavior. Another type, such as the lithium battery management system CN114172233B, while achieving the acquisition and equalization control of basic parameters like voltage and temperature, is limited to the traditional BMS framework. Its control logic is still triggered by preset thresholds, failing to treat the entire vehicle's energy as a unified optimization object and lacking deep fusion of multi-source sensor data and AI-enabled global scheduling capabilities. These shortcomings result in significant deficiencies in the breadth of perception, decision-making intelligence, and scenario adaptability of existing solutions, particularly in handling the complex situation of rapidly changing energy supply and demand and multiple conflicting objectives (such as range, comfort, and battery life) during RV mobility. Therefore, a new system architecture and method are urgently needed that can deeply integrate real-time data from multiple sensors, embed lightweight AI inference capabilities, and achieve end-to-end adaptive energy scheduling. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive energy management system and method for RVs based on AI and multi-sensor fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adaptive energy management method for RVs based on AI and multi-sensor fusion, comprising the following steps:

[0007] The environmental sensing module acquires information about the light intensity, ambient temperature, humidity, wind speed, and geographical location of the RV's external environment.

[0008] The vehicle status monitoring module acquires information such as the RV's driving speed, engine start / stop status, generator output power, chassis power supply status, and parking status.

[0009] The battery health monitoring module acquires the individual cell voltage, total voltage, charge and discharge current, battery surface temperature, internal temperature rise rate, cumulative charge and discharge cycle count, and state of charge value calculated based on the fusion of open-circuit voltage method and ampere-hour integral method of the main energy storage battery pack.

[0010] The user behavior recognition module obtains real-time power consumption of various electrical loads in the vehicle, historical power consumption patterns, user-defined priority strategies, start / stop commands for high-power devices such as air conditioners and water heaters, and explicit input of user energy usage preferences.

[0011] The above four types of data are synchronized and aligned according to a unified time base to form a structured multi-source heterogeneous time series dataset.

[0012] The multi-source heterogeneous time-series dataset is input into a pre-trained lightweight multimodal feature fusion neural network model, which includes a convolutional encoder for extracting low-dimensional representations of the environment and vehicle state, a gated recurrent unit for modeling long-term dependencies of battery health state, an attention mechanism module for parsing user behavior patterns, and a graph neural network layer for cross-modal feature interaction and conflict resolution.

[0013] Based on the output of the lightweight multimodal feature fusion neural network model, the optimal energy dispatch strategy for the next time window is generated. The strategy includes photovoltaic charging power allocation ratio, generator start-stop decision, battery charge-discharge depth limit threshold, power supply guarantee level for high-priority loads, and dynamic power curtailment instructions for non-essential loads.

[0014] The optimal energy dispatch strategy is sent to the energy execution control unit, which drives the photovoltaic controller, generator controller, battery management system and load distribution module to perform corresponding actions to complete closed-loop control.

[0015] In one embodiment of the present invention, the environmental perception module includes a light intensity sensor installed on the roof of the RV, a temperature and humidity composite sensor set on the outside of the vehicle body, a geolocation unit integrated into the navigation system, and a wind speed and direction detection device. All sensors are connected to the central data acquisition unit via a CAN bus or an RS485 communication interface.

[0016] In one embodiment of the present invention, the vehicle status monitoring module obtains the driving speed and engine status by reading the standardized parameters of the vehicle OBD interface, and directly measures the electrical parameters of the generator output terminal through a current transformer and a voltage transmitter. The parking status is determined by a mechanical limit switch or a Hall position sensor.

[0017] As one embodiment of the present invention, the battery health monitoring module adopts a distributed slave control unit architecture. Each slave control unit is responsible for monitoring a group of series-connected battery cells, collecting their voltage and temperature, and uploading the data to the master control unit through an isolated SPI bus. The master control unit performs joint state of charge estimation based on the extended Kalman filter algorithm and calculates the battery internal resistance change rate to assess the degree of aging.

[0018] As one embodiment of the present invention, the user behavior recognition module obtains the real-time active and reactive power of each load through the power metering chip built into the smart socket, and combines it with the power consumption preference configuration file set by the user through the human-computer interaction interface to construct a user power consumption behavior profile, which is updated in units of fixed-length time windows.

[0019] As one embodiment of the present invention, the synchronization alignment of the multi-source heterogeneous time series dataset is achieved by combining hardware trigger signals and software interpolation. The hardware trigger signal is generated by a central clock source and is used to mark the sampling time of each sensor. For data streams with inconsistent sampling frequencies, cubic spline interpolation is used to unify them into a standard time series sequence with twenty sampling points per second.

[0020] As one embodiment of the present invention, the lightweight multimodal feature fusion neural network model is deployed on an embedded AI acceleration chip. The chip supports INT8 quantization inference, the total number of model parameters does not exceed one million, the inference latency is less than fifty milliseconds, and the model training stage adopts a transfer learning strategy. It is first pre-trained on a large-scale general energy scheduling dataset, and then fine-tuned on a small sample dataset of a specific RV operation scenario.

[0021] As one embodiment of the present invention, the graph neural network layer models the four dimensions of environment, vehicle, battery and user as nodes in the graph, and defines the physical coupling relationship and logical dependency relationship between nodes as edges. The weighted aggregation and conflict resolution of cross-domain features are realized through message passing mechanism, and finally the control weight coefficients of each energy execution unit are output.

[0022] As one embodiment of the present invention, after receiving the scheduling strategy, the energy execution control unit first verifies the physical feasibility of the strategy, including the maximum allowable charging and discharging current of the battery, the minimum stable operating power of the generator, and the current maximum power point tracking capability of the photovoltaic array. If the strategy is feasible, it is executed immediately; if it is not feasible, a constraint correction algorithm is started to generate a suboptimal feasible strategy under the premise of satisfying the hard safety boundary.

[0023] According to another aspect of the present invention, an adaptive energy management system for RVs based on AI and multi-sensor fusion is provided, comprising:

[0024] The environmental sensing module is used to acquire information such as light intensity, ambient temperature, humidity, wind speed, and geographical location of the RV's external environment.

[0025] The vehicle status monitoring module is used to acquire information such as the RV's driving speed, engine start / stop status, generator output power, chassis power supply status, and parking status.

[0026] The battery health monitoring module is used to acquire the individual cell voltage, total voltage, charge and discharge current, battery surface temperature, internal temperature rise rate, cumulative charge and discharge cycle count, and state of charge value of the main energy storage battery pack.

[0027] The user behavior recognition module is used to obtain the real-time power consumption of each electrical load in the vehicle, historical power consumption patterns, user-defined priority strategies, start / stop commands for high-power devices, and explicit input of user energy usage preferences.

[0028] The central data fusion and AI inference module is used to synchronize and align the above four types of data and then input them into a lightweight multimodal feature fusion neural network model to generate the optimal energy dispatch strategy.

[0029] The energy execution control unit is used to receive the optimal energy dispatch strategy and drive the photovoltaic controller, generator controller, battery management system and load distribution module to perform corresponding actions.

[0030] As one embodiment of the present invention, the central data fusion and AI inference module is integrated on a system-on-a-chip with a dual-core ARM Cortex-A72 processor and a dedicated NPU coprocessor, running a real-time operating system to ensure deterministic latency in data processing and control command issuance.

[0031] In one embodiment of the present invention, the load distribution module is composed of a solid-state relay array, with each relay corresponding to a power circuit and having overcurrent protection and remote opening and closing functions. Its control signal is issued by the energy execution control unit through an isolated digital output interface.

[0032] As one embodiment of the present invention, the photovoltaic controller adopts the maximum power point tracking algorithm. Its real-time tracking step size is dynamically adjusted according to the predicted value of light change output by the AI ​​inference module. When the light fluctuates drastically, it adopts a small step size for fine tracking and a large step size for fast convergence when the light is stable.

[0033] As one embodiment of the present invention, after receiving the charge / discharge depth limit threshold, the battery management system adjusts the duty cycle of the bidirectional DC-DC converter to control the actual charge / discharge current of the battery within a safe range, and reports the execution status to the central data fusion and AI inference module in real time for the next round of strategy iteration.

[0034] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention constructs a holographic perception system covering four dimensions: environment, vehicle, battery, and user. It breaks through the limitations of traditional energy management systems that rely solely on a single data source such as the meter or BMS, achieving a panoramic, real-time mapping of the RV energy ecosystem. By deploying a lightweight multimodal feature fusion neural network model, this invention transforms multi-source heterogeneous data into physically meaningful control strategies, solving the problem of static control logic and inability to dynamically adapt to complex mobile scenarios in existing technologies. This invention introduces graph neural networks for cross-modal feature interaction and conflict resolution, effectively addressing semantic gaps and logical contradictions between different data sources, ensuring the global consistency and local feasibility of the scheduling strategy. The closed-loop control architecture of this invention directly maps AI inference results to the execution unit and incorporates a built-in physical constraint verification mechanism, ensuring both the intelligence of the control and the safety and robustness of the system. Ultimately, this invention significantly improves the overall efficiency of the RV energy system, reducing ineffective energy consumption by more than 15% under typical operating conditions, extending the cycle life of the lithium-ion battery pack by more than 20%, and providing users with a seamless and personalized energy usage experience. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating the workflow of the adaptive energy management system and method for RVs based on AI and multi-sensor fusion, as described in this invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figure 1This invention provides a technical solution: An adaptive energy management method for RVs based on AI and multi-sensor fusion is offered. Its core lies in acquiring multi-source heterogeneous data on the environment, vehicle, battery, and user behavior through a four-dimensional real-time perception system. It then relies on a lightweight multimodal feature fusion neural network model for online inference to generate the optimal energy scheduling strategy for the next time window. Finally, the energy execution control unit completes closed-loop control. The following will describe in detail the specific implementation steps of this method.

[0038] The method first executes step S101: acquiring information on the light intensity, ambient temperature, humidity, wind speed, and geographical location of the RV's external environment through the environmental sensing module. The environmental sensing module consists of a light intensity sensor installed on the roof of the RV, a temperature and humidity composite sensor located on the exterior of the vehicle, a geolocation unit integrated into the navigation system, and a wind speed and direction detection device. All sensors establish a stable communication link with the central data acquisition unit via a CAN bus or RS485 communication interface. The light intensity sensor employs a silicon photodiode array structure, covering a range of 0 to 120,000 lux, with a sampling frequency of 10 times per second. The temperature and humidity composite sensor integrates a capacitive humidity sensor and a thermistor, with a temperature measurement range of -40°C to +80°C and an accuracy of ±0.5°C, and a humidity measurement range of 0% to 100% relative humidity and an accuracy of ±3%. The geolocation unit is based on a BeiDou / GPS dual-mode receiver, outputting latitude, longitude, altitude, and UTC timestamp, with a positioning accuracy better than 5 meters. The wind speed and direction detection device uses the ultrasonic time-of-flight method, with no moving mechanical parts, measuring wind speed from 0 to 60 meters per second with a resolution of 0.1 meters per second. All of these sensors enter continuous operation mode upon system power-on. Their raw data, after local filtering, is encapsulated in a fixed frame format and periodically uploaded to the central data acquisition unit.

[0039] The following step, S102, is executed: The vehicle status monitoring module acquires information on the RV's driving speed, engine start / stop status, generator output power, chassis power supply status, and parking status. The vehicle status monitoring module obtains driving speed and engine status by reading standardized parameters from the onboard OBD interface. Driving speed is derived from wheel speed signals broadcast via the SAE J1939 protocol data frame through the CAN bus. Engine start / stop status is determined by parsing the operating flag bit issued by the engine control unit. Generator output power is obtained by directly measuring the electrical parameters at the output terminals of a current transformer and a voltage transmitter. The rated transformation ratio of the current transformer is 500 amperes to 1 ampere, and the input range of the voltage transmitter is 0 to 50 volts DC. The output signals of both are isolated and amplified before being sent to an analog-to-digital converter, with a sampling frequency of 100 times per second. The chassis power supply status refers to whether the 12-volt low-voltage system is powered by the engine-driven AC generator. This status is determined by detecting the voltage threshold of the main line of the low-voltage distribution box, which is set to 13.5 volts. The parking status is determined by a mechanical limit switch installed at the end of the handbrake lever or a Hall position sensor embedded in the transmission housing. When the handbrake is fully engaged or the transmission is in P gear, a high-level signal is output to indicate that the parking status is activated. All vehicle status parameters are marked with millisecond-level timestamps and are synchronously transmitted to the central data acquisition unit via the internal CAN network.

[0040] Next, step S103 is executed: the battery health monitoring module acquires the individual cell voltages, total voltages, charge / discharge currents, battery surface temperatures, internal temperature rise rates, cumulative charge / discharge cycle counts, and state-of-charge (SOC) values ​​calculated using a fusion of the open-circuit voltage method and the ampere-hour integral method for the main energy storage battery pack. The battery health monitoring module employs a distributed slave unit architecture. Each slave unit monitors a battery module consisting of sixteen lithium iron phosphate cells connected in series. It acquires the voltage of each individual cell through a sixteen-channel high-precision analog front-end chip, with a voltage measurement range of 2.0 to 4.2 volts and an accuracy of ±1.5 millivolts. Simultaneously, each module is equipped with four NTC thermistors attached to the surface of key cells, providing a temperature sampling accuracy of ±0.5 degrees Celsius. All slave units upload data to the master unit via an isolated SPI bus at a communication rate of 2 MHz, employing cyclic redundancy check (CRC) to ensure data integrity. The main control unit performs joint state-of-charge estimation based on the extended Kalman filter algorithm. Its state equation integrates an open-circuit voltage-state-of-charge lookup table function with a dynamic correction term based on the ampere-hour integral. The process noise covariance matrix is ​​dynamically adjusted according to the battery aging degree. The internal temperature rise rate is obtained by differentiating the temperature sequence through a sliding window with a window length of 30 seconds. The cumulative charge-discharge cycle count is depth-weighted, with each discharge depth exceeding 20% ​​counted as an equivalent cycle. The state-of-charge value is updated once per second and is sent along with other battery parameters.

[0041] Next, step S104 is executed: The user behavior recognition module acquires real-time power consumption, historical power consumption patterns, user-defined priority strategies, start / stop commands for high-power devices such as air conditioners and water heaters, and explicit user input regarding energy usage preferences for each electrical load within the vehicle. The user behavior recognition module obtains the real-time active and reactive power of each load through the power metering chip built into the smart socket. This power metering chip uses a dedicated energy metering integrated circuit, supporting measurements of multiple parameters such as voltage, current, power factor, and harmonic content, with a sampling frequency of 8,000 times per second. The fundamental active power is output after Fourier transform by a digital signal processor. All electrical circuits are equipped with independent smart sockets, with communication interfaces using power line carrier or ZigBee wireless protocols, and data is aggregated to the central gateway. Historical power consumption patterns are statistically analyzed on a 24-hour cycle, recording the average power, peak power, and usage duration for each time period, forming a time-series feature vector. User-defined priority strategies are configured through a human-computer interaction interface, divided into high, medium, and low levels. High-priority loads include lighting, communication equipment, and medical instruments; medium-priority loads include refrigerators and water pumps; and low-priority loads include air conditioners, water heaters, and entertainment systems. Start / stop commands for high-power devices are triggered manually by the user or automatically by a scheduled task. Commands include the target device identifier, action type (on / off), and duration. Explicit user input regarding energy usage preferences includes three mode options: "Energy Saving Priority," "Comfort Priority," and "Fast Charging." These options directly affect the optimization target weight of subsequent scheduling strategies. All user behavior data is reported in an event-driven manner with precise timestamps.

[0042] After completing the acquisition of the above four types of data, step S105 is executed: the four types of data are synchronized and aligned according to a unified time base to form a structured multi-source heterogeneous time-series dataset. Synchronization alignment is achieved using a combination of hardware trigger signals and software interpolation. The central clock source generates a hardware trigger pulse with a period of fifty milliseconds, which is simultaneously distributed to all sensor subsystems as a sampling synchronization signal. For third-party devices that cannot respond to hardware triggers (such as OBD interfaces), their data arrival time is recorded and aligned with the most recent hardware trigger time. Due to the inconsistent original sampling frequencies of the various sensors (10 Hz for environmental sensors, 100 Hz for vehicle status, 1 Hz for battery parameters, and user behavior event-driven data), all data streams need to be unified to a standard time-series sequence of twenty sampling points per second. For low-frequency data (such as battery state of charge), cubic spline interpolation is used to generate intermediate points between adjacent effective values; for high-frequency data (such as vehicle speed), a moving average downsampling to the target frequency is used. The interpolation process is executed in the real-time operating system task of the central data acquisition unit to ensure that the maximum alignment error does not exceed twenty-five milliseconds. The resulting multi-source heterogeneous time-series dataset is stored in the form of a two-dimensional tensor, with the row dimension representing the time step (twenty steps per second) and the column dimension representing the feature channels (a total of one hundred and twenty-eight dimensions). Each element is a floating-point value, representing the normalized value of the corresponding physical quantity at that time step.

[0043] Then, step S106 is executed: the multi-source heterogeneous time-series dataset is input into a pre-trained lightweight multimodal feature fusion neural network model. This model includes a convolutional encoder for extracting low-dimensional representations of the environment and vehicle state, a gated recurrent unit for modeling long-term dependencies in battery health state, an attention mechanism module for parsing user behavior patterns, and a graph neural network layer for cross-modal feature interaction and conflict resolution. This neural network model is deployed on an embedded AI acceleration chip, specifically a system-on-a-chip (SoC) with a dedicated NPU coprocessor, supporting INT8 quantization inference. The model has a total of 980,000 parameters, and the measured inference latency is 42 milliseconds. During the model training phase, a transfer learning strategy is employed. Pre-training is first performed on a large-scale dataset containing 100,000 hours of general mobile energy scheduling data, followed by fine-tuning on a small sample dataset of 5,000 hours specific to RV operation scenarios. The fine-tuning phase uses an Adam optimizer with weight decay, an initial learning rate of 0.001, and a batch size of 64.

[0044] The convolutional encoder consists of three layers of one-dimensional causal convolutions, each with a kernel size of five and channel numbers of 64, 128, and 256 respectively. The activation function is a modified linear unit (MRU), used to extract local temporal patterns from environmental and vehicle state sequences. The gated recurrent unit (GRU) comprises two stacked layers with 512 hidden units, used to capture long-term dependencies of battery health status over a timescale of several hundred seconds. Its input is a subset of battery-related features (32 dimensions). The attention mechanism module employs a multi-head self-attention structure with eight heads, used to analyze key events and temporal relationships in user behavior sequences. Its query, key, and value vectors are mapped from user behavior features (40 dimensions) by fully connected layers. The graph neural network layer models the four dimensions of environment, vehicle, battery, and user as nodes in the graph. Node features are the outputs of the three encoders and the original user behavior embedding vector, respectively. The physical coupling and logical dependencies between nodes are defined as edges, such as twelve predefined edges including "light intensity affects photovoltaic output," "driving status affects generator availability," and "battery state of charge constrains load power supply capacity." The message passing mechanism uses the following formula for feature aggregation: ,in, Indicates the first Layer nodes eigenvectors, Its set of neighboring nodes, These are the learnable edge weight coefficients. To share the weight matrix, This is a non-linear activation function. After three layers of graph convolution, the features of each node are concatenated and fed into the fully connected decoder.

[0045] The decoder consists of a three-layer fully connected network with five output dimensions, corresponding to the photovoltaic charging power allocation ratio, generator start / stop decision, battery charge / discharge depth limit threshold, high-priority load power supply guarantee level, and dynamic power curtailment command for non-essential loads. Specifically, the photovoltaic charging power allocation ratio is a continuous value between zero and one, representing the proportion of currently available photovoltaic power allocated to battery charging; the generator start / stop decision is a binary discrete value, with zero indicating shutdown and one indicating startup; the battery charge / discharge depth limit threshold is a percentage value ranging from 10% to 90%, used to dynamically adjust the available battery capacity window; the high-priority load power supply guarantee level is a Boolean value, ensuring basic functionality is maintained during energy shortages; and the dynamic power curtailment command for non-essential loads is a tiered integer value, ranging from zero (no power curtailment) to three (complete shutdown).

[0046] Execute step S107: Based on the output of the lightweight multimodal feature fusion neural network model, generate the optimal energy scheduling strategy for the next time window. The time window length of this strategy is ten seconds, meaning the model outputs a new set of control commands every ten seconds. During strategy generation, the model output values ​​need to undergo inverse normalization to convert them into actual control parameters in the physical world. For example, the photovoltaic charging power allocation ratio is multiplied by the current maximum power point tracking output power of the photovoltaic array to obtain the actual charging power allocated to the battery; the battery charge / discharge depth limit threshold is compared with the current state of charge. If the current state of charge is below the lower limit of the threshold, discharging is prohibited; if it is above the upper limit, charging is prohibited.

[0047] Finally, step S108 is executed: the optimal energy dispatch strategy is sent to the energy execution control unit, which drives the photovoltaic controller, generator controller, battery management system, and load distribution module to perform corresponding actions, completing closed-loop control. The energy execution control unit first verifies the physical feasibility of the strategy, including the maximum allowable charge and discharge current of the battery (reported in real time by the battery health monitoring module), the minimum stable operating power of the generator (set to 30% of the rated power), and the current maximum power point tracking capability of the photovoltaic array (feedback from the photovoltaic controller). If the strategy is feasible, a control signal is immediately sent through the isolated digital output interface; if it is not feasible, a constraint correction algorithm is initiated to generate a suboptimal feasible strategy under the premise of satisfying the hard safety boundary. The constraint correction algorithm uses the projected gradient descent method to iteratively correct the infeasible solution along the constraint gradient direction until all physical constraints are satisfied.

[0048] The photovoltaic controller employs a maximum power point tracking (MPPT) algorithm, with its real-time tracking step size dynamically adjusted based on the predicted irradiance changes output by the AI ​​inference module. When the predicted rate of change in irradiance exceeds 5,000 lux per second within the next five seconds, the tracking step size is set to 0.5 volts; otherwise, it is set to 2 volts. Upon receiving start / stop commands, the generator controller sends a request signal to the engine control unit via the CAN bus and monitors feedback to confirm the execution status. Upon receiving the charge / discharge depth limit threshold, the battery management system adjusts the duty cycle of the bidirectional DC-DC converter to control the actual charge / discharge current of the battery within a safe range and reports the execution status in real-time to the central data fusion and AI inference module for the next round of strategy iteration. The load distribution module uses a solid-state relay array, with each relay corresponding to a power circuit. It features overcurrent protection and remote tripping functions, and its control signals are sent by the energy execution control unit through an isolated digital output interface, with a response delay of less than ten milliseconds.

[0049] The above-described methods and steps constitute a complete adaptive energy management closed loop, achieving end-to-end automation from multi-source sensing and intelligent reasoning to precise execution. Under typical RV operating conditions, this method can reduce ineffective energy consumption by more than 15%, extend the cycle life of lithium-ion battery packs by more than 20%, and provide users with a seamless and personalized energy usage experience.

[0050] At the system level, this invention also provides an adaptive energy management system for RVs based on AI and multi-sensor fusion. This system includes an environmental perception module, a vehicle status monitoring module, a battery health monitoring module, a user behavior recognition module, a central data fusion and AI inference module, and an energy execution control unit. The environmental perception module, vehicle status monitoring module, battery health monitoring module, and user behavior recognition module respectively perform the data acquisition functions of steps S101 to S104, and their hardware configuration and communication methods have been detailed in the method section. The central data fusion and AI inference module is integrated on a system-on-a-chip (SoC) with a dual-core ARM Cortex-A72 processor and a dedicated NPU coprocessor, running a real-time operating system to ensure deterministic latency in data processing and control command issuance. This module is responsible for executing steps S105 to S107, including data synchronization and alignment, neural network inference, and strategy generation. The energy execution control unit is an independent safety-critical controller, employing a dual-redundant microcontroller architecture. The main control chip is an automotive-grade MCU compliant with ISO-26262 and ASIL-B standards, responsible for strategy verification and execution driving in step S108. The load distribution module, photovoltaic controller, generator controller and battery management system are all existing mature equipment, but their control interfaces have been adapted to the communication protocol of this system to ensure command compatibility and execution reliability.

[0051] The entire system employs a rigorous timing scheduling mechanism to ensure the coordinated operation of all modules. The central data fusion and AI inference module executes its main loop with a 50-millisecond cycle, with the first 25 milliseconds used for data acquisition and synchronization, the middle 15 milliseconds for neural network inference, and the last 10 milliseconds for strategy issuance and status feedback. The energy execution control unit scans control commands with a 10-millisecond cycle and monitors the actuator status in real time. Once an anomaly is detected (such as relay sticking or current exceeding limits), a safety protection mechanism is immediately triggered, cutting off the relevant circuits and reporting a fault code. The system also has a built-in data logging module that writes key operating parameters to non-volatile memory once per minute for subsequent performance analysis and model iteration optimization.

[0052] In summary, this embodiment fully discloses the technical solution of the present invention by describing each step of the method and its corresponding system implementation in detail, thus meeting the requirement of full disclosure and providing a solid foundation for defining the scope of patent protection.

Claims

1. A method for adaptive energy management of recreational vehicles based on AI and multi-sensor fusion, characterized in that, include: S101: Obtain information on the light intensity, ambient temperature, humidity, wind speed, and geographical location of the RV's external environment through the environmental perception module; S102: Obtain information on the RV's driving speed, engine start / stop status, generator output power, chassis power supply status, and parking status through the vehicle status monitoring module; S103: The battery health monitoring module acquires the individual cell voltage, total voltage, charge and discharge current, battery surface temperature, internal temperature rise rate, cumulative charge and discharge cycle count, and state of charge value calculated based on the fusion of open-circuit voltage method and ampere-hour integral method of the main energy storage battery pack. S104: The user behavior recognition module obtains the real-time power consumption of each electrical load in the vehicle, historical power consumption patterns, user-defined priority strategies, start / stop commands of high-power devices such as air conditioners and water heaters, and explicit input of user energy usage preferences. S105: Synchronize and align the above four types of data according to a unified time base to form a structured multi-source heterogeneous time series dataset; S106: Input the multi-source heterogeneous time-series dataset into a pre-trained lightweight multimodal feature fusion neural network model. The model includes a convolutional encoder for extracting low-dimensional representations of the environment and vehicle state, a gated recurrent unit for modeling long-term dependencies of battery health state, an attention mechanism module for parsing user behavior patterns, and a graph neural network layer for cross-modal feature interaction and conflict resolution. S107: Based on the output of the lightweight multimodal feature fusion neural network model, generate the optimal energy scheduling strategy in the next time window. The strategy includes photovoltaic charging power allocation ratio, generator start-stop decision, battery charge-discharge depth limit threshold, power supply guarantee level for high-priority loads, and dynamic power limiting instructions for non-essential loads. S108: The optimal energy dispatch strategy is sent to the energy execution control unit, which drives the photovoltaic controller, generator controller, battery management system and load distribution module to perform corresponding actions to complete closed-loop control.

2. The adaptive energy management method for RVs based on AI and multi-sensor fusion as described in claim 1, characterized in that, The above four types of data are synchronized and aligned according to a unified time base to form a structured multi-source heterogeneous time-series dataset, including: The hardware trigger signal generated by the central clock source is used as the sampling synchronization reference to timestamp the raw data of each sensor subsystem. For data streams with inconsistent sampling frequencies, cubic spline interpolation or moving average downsampling is used to unify them into a standard time sequence of 20 sampling points per second. The interpolated data is stored in the form of a two-dimensional tensor, with the row dimension representing the time step and the column dimension representing the feature channel. Each element is the normalized value of the corresponding physical quantity at that time step.

3. The adaptive energy management method for RVs based on AI and multi-sensor fusion as described in claim 2, characterized in that, The process of inputting the multi-source heterogeneous time-series dataset into a pre-trained lightweight multimodal feature fusion neural network model includes: A subset of environmental and vehicle state-related features is input into a convolutional encoder, and local temporal patterns are extracted through three layers of one-dimensional causal convolution to obtain a low-dimensional representation vector of the environment and vehicle. A subset of battery health-related features is input into a gated recurrent unit, and long-term dependencies are modeled through a two-layer stacked structure to obtain a battery health state temporal encoding vector. A subset of user behavior-related features is input into the attention mechanism module, and the key events and temporal correlations are analyzed through a multi-head self-attention structure to obtain the user behavior attention embedding vector. The environment-vehicle low-dimensional representation vector, battery health status temporal encoding vector, user behavior attention embedding vector, and original user behavior embedding vector are used as graph node features to construct a heterogeneous graph containing four nodes: environment, vehicle, battery, and user.

4. The adaptive energy management method for RVs based on AI and multi-sensor fusion according to claim 3, characterized in that, The graph neural network layer achieves cross-modal feature interaction and conflict resolution through a message passing mechanism, including: Based on predefined physical coupling and logical dependency relationships, an edge set is constructed between nodes. The edge set includes twelve edges: "sunlight intensity affects photovoltaic output", "driving status affects generator availability", and "battery state of charge constrains load power supply capacity". In each graph convolutional layer, the neighbor features of each node are weighted and aggregated, and the weight coefficients are determined by the learnable edge weight parameters. After three layers of graph convolution, the final feature vectors of each node are concatenated and input into a fully connected decoder, which outputs a five-dimensional control instruction vector.

5. The adaptive energy management method for RVs based on AI and multi-sensor fusion according to claim 4, characterized in that, Based on the output of the lightweight multimodal feature fusion neural network model, the optimal energy scheduling strategy for the next time window is generated, including: The five-dimensional control command vector output by the decoder is denormalized and converted into actual control parameters in the physical world. Multiply the photovoltaic charging power allocation ratio by the current maximum power point tracking output power of the photovoltaic array to obtain the actual charging power allocated to the battery; The battery charge / discharge depth limit threshold is compared with the current state of charge. If the current state of charge is lower than the lower limit of the threshold, discharging is prohibited; if it is higher than the upper limit, charging is prohibited. Dynamic power limiting commands for non-essential loads are mapped to tiered integer values, ranging from zero (no power limiting) to three (complete disconnection of the corresponding load circuit).

6. The adaptive energy management method for RVs based on AI and multi-sensor fusion according to claim 5, characterized in that, The optimal energy dispatch strategy is sent to the energy execution control unit, which then drives the photovoltaic controller, generator controller, battery management system, and load distribution module to perform corresponding actions, including: The energy execution control unit first verifies the physical feasibility of the strategy, which includes the maximum allowable charge and discharge current of the battery, the minimum stable operating power of the generator, and the current maximum power point tracking capability of the photovoltaic array. If the strategy is feasible, immediately send a control signal through the isolated digital output interface; If the strategy is not feasible, the constraint correction algorithm is initiated to generate a suboptimal feasible strategy under the premise of satisfying the hard safety boundary. The constraint correction algorithm uses the projected gradient descent method to iteratively correct the infeasible solution along the constraint gradient direction.

7. The adaptive energy management method for RVs based on AI and multi-sensor fusion according to claim 6, characterized in that, The photovoltaic controller employs a maximum power point tracking algorithm, and its real-time tracking step size is dynamically adjusted based on the predicted illumination change value output by the AI ​​inference module, including: When the predicted rate of change in light intensity within the next five seconds exceeds 5,000 lux per second, the tracking step size is set to 0.5 volts. When the predicted rate of change in light intensity does not exceed 5,000 lux per second, the tracking step size is set to 2 volts.

8. The adaptive energy management method for RVs based on AI and multi-sensor fusion according to claim 7, characterized in that, After receiving the charge / discharge depth limit threshold, the battery management system adjusts the duty cycle of the bidirectional DC-DC converter to control the actual charge / discharge current of the battery within a safe range, and reports the execution status to the central data fusion and AI inference module in real time for the next round of strategy iteration.

9. A recreational vehicle adaptive energy management system based on AI and multi-sensor fusion, characterized in that, include: The environmental sensing module is used to acquire information such as light intensity, ambient temperature, humidity, wind speed, and geographical location of the RV's external environment. The vehicle status monitoring module is used to acquire information such as the RV's driving speed, engine start / stop status, generator output power, chassis power supply status, and parking status. The battery health monitoring module is used to acquire the individual cell voltage, total voltage, charge and discharge current, battery surface temperature, internal temperature rise rate, cumulative charge and discharge cycle count, and state of charge value of the main energy storage battery pack. The user behavior recognition module is used to obtain the real-time power consumption of each electrical load in the vehicle, historical power consumption patterns, user-defined priority strategies, start / stop commands for high-power devices, and explicit input of user energy usage preferences. The central data fusion and AI inference module is used to synchronize and align the above four types of data and then input them into a lightweight multimodal feature fusion neural network model to generate the optimal energy dispatch strategy. The energy execution control unit is used to receive the optimal energy dispatch strategy and drive the photovoltaic controller, generator controller, battery management system and load distribution module to perform corresponding actions.

10. The RV adaptive energy management system based on AI and multi-sensor fusion according to claim 9, characterized in that, The central data fusion and AI inference module is integrated on a system-on-a-chip with a dual-core ARM Cortex-A72 processor and a dedicated NPU coprocessor, running a real-time operating system to ensure deterministic latency in data processing and control command issuance; the load distribution module is composed of a solid-state relay array, with each relay corresponding to a power circuit, and has overcurrent protection and remote switching functions. Its control signals are issued by the energy execution control unit through an isolated digital output interface.

Images