A wind-solar-storage micro-grid multi-source data prediction and scheduling method for waterless public toilets

Abstract

The application relates to the technical field of micro-grid energy management, and discloses a wind-solar-storage micro-grid multi-source data prediction scheduling method for waterless public toilets, which acquires multi-source operation data, scene characteristic vectors and environmental parameters of the public toilets, and performs physical correction on power consumption load data, photovoltaic power generation data and wind power generation data according to the environmental parameters; the corrected data, SOC data, battery temperature data, meteorological data and the scene characteristic vectors are input into a pre-trained VMD-BiLSTM multi-branch collaborative prediction model to output load, photovoltaic output, wind power output, SOC trajectory and battery temperature prediction values in the next 24 hours; the endurance days are determined according to the prediction results and weather forecast data, and a hierarchical discharge planning is formulated according to the SOC data and the endurance days. The application improves the prediction accuracy by fusing the environmental parameter correction and the multi-branch prediction model, and guarantees the power supply reliability of the public toilets under off-grid conditions through the hierarchical discharge planning.

Description

Technical Field

[0001] This application relates to the field of microgrid energy management technology, and in particular to a multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets. Background Technology

[0002] With the advancement of ecological civilization construction and the green development of infrastructure, the application of waterless ecological toilets (i.e., waterless public toilets) is increasing in areas where conventional municipal water supply and drainage networks are difficult to cover, such as plateaus, Gobi deserts, wastelands, and remote scenic areas. These public toilets do not require flushing and rely on microbial degradation or incineration to treat excrement, significantly saving water resources. However, their normal operation still heavily depends on electricity supply, and the load includes lighting, ventilation systems, temperature control heating of biological treatment units, information display screens, etc., with electricity load showing obvious diurnal differences and seasonal fluctuations.

[0003] Currently, some technical solutions propose equipping waterless public toilets with photovoltaic (PV) power generation, wind power generation, and energy storage devices to achieve self-sufficiency in electricity. For example, an integrated PV-storage power supply system can provide electricity to the toilets, or a wind-solar hybrid power generation method can be used. However, while plateau and Gobi desert regions are rich in solar and wind energy resources, their output is significantly intermittent and uncertain. Sunlight intensity fluctuates drastically with cloud cover changes, and wind speeds and directions are unpredictable, making it difficult to maintain stable PV and wind power generation. When power generation exceeds the electricity load, surplus energy needs to be stored in energy storage batteries; when power generation is insufficient, the energy storage batteries need to release energy to supplement it. Therefore, maintaining a dynamic balance between power generation and energy storage is a crucial prerequisite for waterless public toilets to achieve reliable power supply in such scenarios.

[0004] Existing solutions often rely on simple combinations of photovoltaic panels, wind turbines, and batteries. Their control logic is typically rudimentary, lacking systematic prediction and optimized scheduling of toilet power load, photovoltaic and wind power output, and energy storage status over a prolonged period. In extreme weather conditions such as continuous rain or no wind, these passively responsive systems struggle to guarantee reliable and continuous power supply. Particularly in waterless public toilets, the biodegradation unit requires a certain heating temperature to ensure normal reaction. If a power outage occurs due to SOC prediction errors or improper scheduling, the degradation process will halt. Even after power is restored, the biodegradation reaction takes several hours to re-establish, rendering the toilet unusable. Therefore, accurate prediction of SOC changes and rational allocation of power accordingly are crucial for ensuring the continuous operation of public toilets. There is an urgent need for a multi-source data prediction and scheduling method for wind-solar-storage microgrids specifically tailored for waterless public toilet scenarios, meeting the comprehensive requirements of waterless public toilets for power supply reliability, operational economy, and long equipment lifespan in complex and variable environments. Summary of the Invention

[0005] (a) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets, in order to solve the following problems existing in the prior art: (1) Existing public toilet power supply schemes are mostly simple combinations of photovoltaic panels, wind turbines and batteries. The control logic is crude and lacks systematic prediction and optimization scheduling of future power load, wind and solar power output and energy storage status, resulting in technical problems such as low utilization rate of new energy and poor power supply reliability. (2) General microgrid prediction methods are not customized for public toilet scenarios and cannot adapt to the special characteristics of public toilets, such as large fluctuations in electricity load, significant seasonal differences, and difficulty in continuously accumulating historical data due to mobile deployment. This leads to the technical problem of decreased accuracy of prediction models after migration. (3) Existing off-grid energy management schemes lack a forward-looking scheduling mechanism based on prediction. Under completely off-grid conditions, they cannot dynamically estimate the number of days of continuous operation based on the energy storage status and meteorological conditions and actively formulate graded discharge plans, resulting in technical problems such as premature depletion of energy storage and power outages in public toilets.

[0006] (II) Technical Solution To achieve the above objectives, the main technical solutions adopted in this application include: This application provides a method for multi-source data prediction and scheduling of a wind-solar-storage microgrid for waterless public toilets, the specific steps of which include: S1, acquire multi-source operation data, scene feature vector and environmental parameters of the target public toilet. The multi-source operation data includes hourly electricity load data, hourly photovoltaic power generation data, hourly wind power generation data, SOC data, energy storage battery temperature data and meteorological data. S2, the hourly electricity load data, hourly photovoltaic power generation data and hourly wind power generation data are corrected according to the environmental parameters to obtain the corrected corresponding data; S3, input the corrected corresponding data, SOC data, energy storage battery temperature data, meteorological data, and scene feature vector into the pre-trained VMD-BiLSTM multi-branch collaborative prediction model, and output the prediction results. The prediction results include hourly electricity load predictions, hourly photovoltaic power generation predictions, hourly wind power generation predictions, SOC trajectory predictions, and energy storage battery temperature predictions for the next 24 hours. The VMD-BiLSTM multi-branch collaborative prediction model includes a VMD decomposition layer, a shared BiLSTM encoder, and a multi-branch decoder. S4. Determine the remaining power capacity based on the prediction results and the pre-acquired weather forecast data, and formulate a graded discharge plan based on the SOC data and the remaining power capacity. The graded discharge plan divides the remaining power capacity into multiple power capacity levels, and each power capacity level corresponds to a preset load reduction ratio.

[0007] Optionally, in some embodiments of this application, the training process of the VMD-BiLSTM multi-branch collaborative prediction model includes: S01, Obtain the training dataset, which includes hourly electricity load data, hourly photovoltaic power generation data, hourly wind power generation data, SOC data, energy storage battery temperature data, meteorological data, scene feature vectors, and the true value labels corresponding to each prediction branch. S02, Construct an initial VMD-BiLSTM multi-branch collaborative prediction model, which includes a VMD decomposition layer, a shared BiLSTM encoder, and a multi-branch decoder; the multi-branch decoder includes an electricity load prediction branch, a photovoltaic power generation prediction branch, a wind power generation prediction branch, a SOC trajectory prediction branch, and an energy storage battery temperature prediction branch. S03, the training dataset is processed sequentially by the VMD decomposition layer and the shared BiLSTM encoder, and then input into the multi-branch decoder for prediction to obtain the prediction values ​​of each branch; S04, Construct the total loss function that includes the branch prediction loss term and the additional constraint term; S05, with the goal of minimizing the total loss function, the parameters of the initial VMD-BiLSTM multi-branch collaborative prediction model are iteratively optimized using a preset training strategy until the preset convergence condition is met, thus obtaining the trained VMD-BiLSTM multi-branch collaborative prediction model; wherein, during the parameter iterative optimization process, the parameters of the shared BiLSTM encoder adopt a first learning rate, and the parameters of each branch of the multi-branch decoder adopt a second learning rate, wherein the first learning rate is lower than the second learning rate.

[0008] Optionally, in some embodiments of this application, the branch prediction loss term is a weighted sum of the loss functions of each branch, and the initial value of the loss weight of each branch is determined according to the preset safety cost level of the corresponding branch. The safety cost level is used to characterize the degree of influence of the prediction deviation of the corresponding branch on the energy supply reliability and equipment safety of the waterless public toilet; wherein, the safety cost level of the electricity load prediction branch and the SOC trajectory prediction branch is higher than that of the photovoltaic power generation prediction branch, the wind power generation prediction branch and the energy storage battery temperature prediction branch. During training, every preset training round, the loss weight is dynamically adjusted based on the rate of change of prediction error of each branch in the most recent verification period; the dynamic adjustment range of the loss weight is constrained by a lower limit, which is determined by the minimum guarantee weight corresponding to the security cost level of each branch.

[0009] Optionally, in some embodiments of this application, the additional constraint term includes a physical constraint regularization term, which specifically includes: The SOC dynamic constraint regularization term is used to apply a dynamic consistency constraint to the SOC trajectory prediction output by the SOC trajectory prediction branch. A penalty value is generated when the SOC change in adjacent time periods in the predicted SOC trajectory exceeds the upper limit of SOC change determined by the maximum charge and discharge rate of the energy storage battery. The photovoltaic power generation upper bound constraint regularization term is used to constrain the physical upper bound of the hourly photovoltaic power generation prediction value output by the photovoltaic power generation prediction branch, and generates a penalty value when the hourly photovoltaic power generation prediction value exceeds the theoretical maximum photovoltaic power generation calculated based on the light intensity in the meteorological data. The energy conservation constraint regularization term is used to constrain the physical consistency of energy conversion. When the change in SOC and the charging / discharging power in adjacent time periods do not meet the preset energy conversion efficiency relationship of the energy storage battery in the SOC trajectory prediction, a penalty value is generated. The charging / discharging power is calculated based on the hourly electricity load prediction value, the hourly photovoltaic power generation prediction value, and the hourly wind power generation prediction value.

[0010] Optionally, in some embodiments of this application, in step S05, the preset training strategy specifically includes a pre-training phase and a fine-tuning phase: During the pre-training phase, the basic parameters of the shared BiLSTM encoder and each prediction branch are trained based on the training dataset, enabling the VMD-BiLSTM multi-branch collaborative prediction model to learn the cross-scenario common temporal features of the target public toilet wind-solar-storage microgrid. During the fine-tuning phase, when the target public toilet is relocated to the new deployment site, the following sub-steps are performed: S051, Freeze the high-level parameters of the shared BiLSTM encoder, and only fine-tune the parameters of each branch in the multi-branch decoder and the low-level parameters of the shared BiLSTM encoder. The training data used for fine-tuning training is the operation data of the target public toilet collected at the new deployment point. S052, when the cumulative operating data duration of the target public toilet at the newly deployed site reaches the preset data duration threshold, the parameter freeze is lifted and all parameters of the shared BiLSTM encoder and the multi-branch decoder are jointly fine-tuned.

[0011] Optionally, in some embodiments of this application, the number of modal layers and the penalty factor of the VMD decomposition layer in step S02 are determined by the following pre-calibration sub-step: Power spectral density estimation was performed on the historical hourly electricity load data, hourly photovoltaic power generation data and hourly wind power generation data in the training dataset, and the number of significant peaks and the bandwidth of each significant peak were extracted for each data point. The modal layer number search space is determined based on the number of significant peaks, and the penalty factor search space is determined based on the bandwidth of each significant peak. Within the combined search domain formed by the modal layer number search space and the penalty factor search space, the VMD decomposition and reconstruction error is minimized as the objective, where the VMD decomposition and reconstruction error is the mean square error between the reconstructed signal and the original signal. The optimal combination of modal layer number and optimal penalty factor is searched using a Bayesian optimization algorithm and configured in the VMD decomposition layer.

[0012] Optionally, in some embodiments of this application, step S2, which involves correcting the hourly electricity load data, hourly photovoltaic power generation data, and hourly wind power generation data based on the environmental parameters to obtain the corrected corresponding data, specifically includes: When the altitude in the environmental parameters is higher than a preset high-altitude threshold, ventilation compensation load correction is performed. The ventilation compensation load correction includes the following sub-steps: S21, based on the altitude, query the preset altitude-minimum safe air change rate mapping table to determine the target air change rate corresponding to the current deployment point; wherein, the altitude-minimum safe air change rate mapping table is obtained based on the minimum ventilation volume required to maintain the oxygen concentration inside the public toilet at different altitudes not lower than the preset safe concentration threshold, and the target air change rate increases non-linearly with the altitude. S22, multiply the difference between the target air exchange rate and the preset benchmark air exchange rate by the unit air exchange power of the public toilet exhaust system to obtain the ventilation compensation load increment; wherein, the unit air exchange power is determined according to the actual air density at the current deployment point; S23, the ventilation compensation load increment is superimposed on each time period of the hourly electricity load data to obtain the corrected hourly electricity load data, and the ventilation compensation load increment is marked as a non-reducible safe load. The non-reducible safe load mark is used to reserve it as the highest priority in the graded discharge plan.

[0013] Optionally, in some embodiments of this application, determining the number of subsequent flights based on the prediction results and the pre-acquired weather forecast data specifically includes: The prediction results also include the confidence interval of the SOC trajectory prediction value, and the uncertainty interval width is obtained based on the confidence interval; Based on the prediction results, calculate the net energy value for each period in the next 24 hours and determine whether there is a net energy gap. When there is a net energy gap, divide the difference between the current SOC data and the pre-acquired lower limit of SOC by the daily average value of the net energy gap to obtain the initial estimate of the number of days of continued flight. Based on the forecast of consecutive low-irradiance days and / or consecutive low-wind-speed days in the weather forecast data, the initial estimate of the number of days of continued air travel is reduced by meteorological risk; the meteorological reduction result is further reduced based on the uncertainty interval width of the SOC trajectory prediction to obtain the number of days of continued air travel. The larger the uncertainty interval width, the smaller the reduction coefficient.

[0014] Optionally, in some embodiments of this application, step S4, which involves formulating a graded discharge plan based on the SOC data and the remaining lifespan, specifically includes: The capacity decay ratio is determined based on the pre-acquired SOH data. The SOC trajectory prediction is then corrected by capacity conversion based on the capacity decay ratio to obtain the capacity-corrected SOC trajectory. Based on the lowest predicted point of the capacity-corrected SOC trajectory in the next 24 hours and the remaining flight time, multiple flight duration levels are generated. When the SOC data decreases sequentially to the trigger threshold corresponding to each battery life level, a preset load reduction is performed according to the corresponding battery life level until only the highest priority load is retained.

[0015] Optionally, in some embodiments of this application, in the graded discharge plan, the preset lower limit of SOC is dynamically adjusted according to the current ambient temperature: When the ambient temperature is lower than the preset low temperature threshold, the SOC lower limit is adjusted to the preset low temperature protection value so as to reserve the energy required for self-heating in the energy storage battery in a low temperature environment.

[0016] (III) Beneficial Effects This application presents a multi-source data prediction and scheduling method for wind, solar and energy storage microgrids for waterless public toilets. It physically corrects load, photovoltaic and wind power data through environmental parameters, eliminates the interference of environmental factors such as low air pressure at high altitudes, dust, and extreme temperature differences on the original data, and improves prediction accuracy from the data source. The VMD-BiLSTM multi-branch collaborative prediction model uses VMD decomposition to decompose non-stationary time series data into stationary components to reduce learning difficulty. It uses a shared BiLSTM encoder to reduce the number of parameters and reduce the risk of overfitting. The multi-branch decoder outputs five predicted values: load, photovoltaic, wind power, SOC trajectory and battery temperature. This allows each task to share underlying features while retaining independent adaptation parameters, thus improving prediction accuracy under conditions of few samples and cross-scene conditions. By fusing scene feature vectors with operational data as input, the model can perceive the differences in the electricity consumption structure and seasonal variation patterns of public toilets. When migrating and deploying, only the scene labels need to be adjusted to adapt to the new scene without retraining, thus solving the problems of difficulty in accumulating historical data of mobile public toilets and the decrease in prediction accuracy after migration. By determining the remaining power capacity based on forecast results and weather forecasts, and combining SOC data to formulate a tiered discharge plan, the system can predict the number of days of operation that can be sustained under off-grid conditions based on SOC trajectory prediction and weather forecasts. It can proactively divide the system into multiple power capacity levels and set corresponding preset load reduction ratios, making the discharge process more refined and orderly. This avoids power outages and shutdowns of public toilets due to premature depletion of energy storage, and ensures the reliability of power supply and basic sanitation service capabilities under off-grid conditions. Attached Figure Description

[0017] Figure 1 The flowchart of the multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets is shown in this application. Figure 2 This is a flowchart of the training steps of the VMD-BiLSTM multi-branch collaborative prediction model in one embodiment of this application. Detailed Implementation

[0018] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0019] Existing power supply management solutions for waterless public toilets mainly suffer from the following three shortcomings: The first category is a simple power supply solution at the hardware level. This type of solution supplies power to public toilets through a simple combination of photovoltaic panels, wind turbines, and batteries. Its control logic is usually quite crude, limited to a passive response mode of "charging when there is light and discharging when there is no light," lacking systematic prediction and optimized scheduling of future electricity load, wind and solar output, and energy storage status. Although plateau and Gobi regions are rich in solar and wind energy resources, their output is significantly intermittent and uncertain, making it difficult to maintain a stable power generation. When encountering extreme weather such as continuous rain or no wind, this type of passive response system cannot effectively guarantee the reliability and continuity of power supply.

[0020] The second category is microgrid energy management methods for general scenarios. These methods typically train predictive models based on long-term, stable historical load data, and then optimize scheduling to reduce operating costs. However, these methods are mainly geared towards conventional industrial, commercial, or residential microgrids and are not customized for scenarios involving waterless public toilets. Waterless public toilets are usually deployed in areas with weak grid coverage and may be relocated due to changes in scenic area planning. Electricity consumption data is characterized by difficulties in accumulation, high volatility, and variable scenarios, making traditional predictive models difficult to apply.

[0021] The third category is experience-based scheduling methods that rely on weather forecasts. These methods acquire weather forecast data and manually or semi-automatically adjust power supply strategies based on predicted wind speed and solar radiation intensity. However, these methods only use weather forecasts as a reference for scheduling decisions, relying on human experience and lacking quantitative forecasts of wind and solar power output and load demand. Furthermore, deviations between weather forecasts and actual conditions are propagated through various levels of the scheduling process, and when forecast deviations are significant, they can mislead scheduling, causing the system to malfunction as expected.

[0022] To address the aforementioned shortcomings, this application provides a multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets. Its core concept is as follows: First, load, photovoltaic, and wind power data are physically corrected using environmental parameters to eliminate interference from environmental factors such as low air pressure at high altitudes, dust storms, and extreme temperature differences. Second, the corrected data, along with SOC data, battery temperature data, meteorological data, and scene feature vectors, are input into a VMD-BiLSTM multi-branch collaborative prediction model. VMD decomposition is used to decompose non-stationary time-series data into stationary components to reduce prediction difficulty. A multi-branch decoder simultaneously outputs five predicted values: load, photovoltaic, wind power, SOC trajectory, and battery temperature. Finally, the remaining power supply is determined based on the prediction results and weather forecast data, and a tiered discharge plan is formulated in conjunction with SOC data, achieving a complete closed loop from forward-looking prediction to active control under off-grid conditions.

[0023] To better explain and facilitate understanding of this application, a detailed description of its embodiments is provided below in conjunction with the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a clearer and more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0024] Example 1: Figure 1This is a flowchart illustrating a multi-source data prediction and scheduling method for a wind-solar-storage microgrid for waterless public toilets, according to this application. This method can be applied to the energy management controller (EMS) of portable waterless public toilets deployed in scenic areas or remote regions with weak grid coverage, such as plateaus, Gobi deserts, and arid regions. It aims to address issues such as extreme environmental parameters (low air pressure, sandstorms, temperature differences), scenario relocation due to mobile deployment, and the difficulty in ensuring power supply reliability under off-grid conditions. Furthermore, in waterless public toilets, excrement relies on microorganisms for biodegradation at suitable temperatures. If the heating device is prematurely shut down due to SOC prediction errors, the degradation temperature drops, and microbial activity decreases accordingly. If the power outage lasts for a prolonged period, the degradation process may completely stop, and the public toilet will lose its processing capacity. Even after power is restored, the biodegradation reaction requires several hours to re-establish stable processing capacity. Therefore, accurately predicting SOC changes and rationally allocating power accordingly to ensure continuous power supply to critical loads under off-grid conditions is one of the core problems this solution aims to solve.

[0025] The method described in this application is applicable to both deployment scenarios with and without external grid access. When the public toilet is deployed in an area with grid power access (such as near a tourist center in a scenic area), the system can use grid power as a supplementary power source to reduce electricity costs through grid-connected operation. When the public toilet is deployed in remote scenic spots or deep in the Gobi Desert, where the grid is not covered, or in scenarios where it is difficult to run power lines, the system relies entirely on wind, solar, and energy storage power supply to ensure the continuous operation of the public toilet. The following explanation uses an off-grid deployment scenario without external grid access as an example.

[0026] In this application scenario, the target public toilet is equipped with a photovoltaic array and a vertical axis fan on its roof, along with a lithium iron phosphate energy storage battery module and an energy management controller. When the public toilet is relocated to a new deployment site or put into operation, the energy management controller automatically acquires multi-source operational data of the target public toilet (including hourly electricity load data, hourly photovoltaic power generation data, hourly wind power generation data, SOC data, energy storage battery temperature data, and meteorological data), scene feature vectors, and environmental parameters, and then... Figure 1 The process shown completes data correction, prediction, and scheduling decisions, and sends scheduling instructions to the energy storage converter and load control switch to ensure the power supply reliability and continuous operation capability of the public toilet under off-grid conditions. The following section combines... Figure 1 The technical solutions provided in the embodiments of this application will be described in detail.

[0027] S1. Acquire multi-source operational data, scene feature vectors, and environmental parameters of the target public toilet. Multi-source operational data includes hourly electricity load data, hourly photovoltaic power generation data, hourly wind power generation data, state of charge (SOC) data of the energy storage battery module, energy storage battery temperature data, and meteorological data. The scene feature vector includes region type labels, season labels, and time period labels. Multi-source operational data is acquired in real time through a sensor network and local monitoring system deployed in the target public restrooms, specifically including: Hourly power load data: Collected by smart meters installed at the entrance of the public toilet's power distribution box, the total hourly power consumption of all electrical equipment in the public toilet, including lighting, ventilation system, biological treatment unit heating, information display screen, etc., is recorded in kW.

[0028] Hourly photovoltaic power generation data: Collected and recorded by the monitoring module built into the photovoltaic inverter, the actual output power of the photovoltaic array at each moment is measured in kW.

[0029] Hourly wind power generation data: Collected and recorded by the monitoring module built into the wind turbine inverter, the actual output power of the vertical axis wind turbine at each moment is measured in kW.

[0030] SOC data of energy storage battery module: Monitored and reported in real time by battery management system (BMS), recording the remaining percentage of energy storage battery module at the current moment (0%~100%).

[0031] Energy storage battery temperature data: The operating temperature of the energy storage battery module at the current moment is collected and recorded by the temperature sensor built into the BMS, and the unit is ℃.

[0032] Meteorological data: collected through a miniature weather station deployed on the roof of the public toilet, including at least ambient temperature (°C) and light intensity (W / m²). 2 ) and wind speed (m / s).

[0033] The collection period for the above data can be set according to actual needs. In this embodiment, it is preferred to collect data once per hour to form an hourly data sequence.

[0034] The scene feature vector is used to identify the current scene of the target public toilet, providing scene perception capabilities for the prediction model. In this embodiment, the scene feature vector consists of the following three labels: Region Type Label: Based on the climate zone of the current deployment location of the target public toilet, the value is either "Severe Cold Region" or "Non-Severe Cold Region". This label is set by the maintenance personnel through the EMS human-machine interface after the public toilet is relocated and deployed, or it is automatically matched based on the GPS coordinates of the deployment location combined with the geographic information system.

[0035] Seasonal tag: Based on the season of the target public toilet's current operating date, the value can be spring, summer, autumn, or winter. This tag is automatically determined by EMS based on the system clock.

[0036] Time Period Label: Based on the time period to which the current time belongs, the value is either "Daytime High Load Period" or "Nighttime Low Load Period". This label is automatically determined by EMS based on a preset time interval, for example, dividing the daytime high load period into 6:00 to 22:00 and the nighttime low load period into 22:00 to 6:00 the next day.

[0037] The regional and seasonal tags together determine the electricity consumption structure of public restrooms. For example, in frigid regions, there is a higher underfloor heating / heating load in winter, while in non-frigid regions, there is a higher air conditioning / cooling load in summer. The time-of-use tag reflects the daily cyclical variation in the frequency of public restroom use. During the day, there is a large flow of visitors and a high rate of equipment activation, while at night, the flow is small and only the basic operating load needs to be maintained.

[0038] Environmental parameters are used to physically correct multi-source operational data. These parameters include the current deployment site's altitude, atmospheric pressure, ambient temperature, dust accumulation on the photovoltaic panel surface, and snow cover. These environmental parameters can be collected in real time by devices such as micro weather stations, barometers, and image sensors.

[0039] The collected raw data is preprocessed by data cleaning, missing value imputation, time alignment and normalization to obtain a standard format multi-source running dataset, which is used for physical correction processing in the subsequent S2 step.

[0040] S2, based on the environmental parameters of the current deployment point, corrects the hourly electricity load data, hourly photovoltaic power generation data, and hourly wind power generation data to obtain the corrected corresponding data.

[0041] The environmental parameters include altitude, air density, dust accumulation on the photovoltaic panel surface, and ambient temperature, which are used to correct physical deviations in the three types of data, respectively.

[0042] For hourly wind power generation data, corrections are made based on the altitude of the current deployment site and the actual air density. In high-altitude areas, the air density is lower, and the actual output power at the same wind speed is lower than the rated power at the standard air density. This is corrected by the air density ratio.

[0043] In a preferred embodiment, portable toilets are often deployed in areas with abundant wind resources but harsh climates, such as plateaus and deserts. Their power supply is highly dependent on wind power generation, making reliable operation of the wind turbines in low-temperature environments crucial. Therefore, in the low-temperature winter environment of high-altitude scenic areas, wind turbine blades are prone to icing or snow accumulation, leading to deterioration of the wind turbine's aerodynamic characteristics and a significant decrease in power generation efficiency. Even air density correction cannot compensate for the output loss. To address these issues, this embodiment adds a low-temperature anti-icing and de-icing mechanism to the wind turbine control strategy: when the ambient temperature is below a preset icing threshold (preferably -5℃ in this embodiment) and the wind turbine is in a stopped state, the EMS executes an anti-icing pre-start procedure before starting the wind turbine. This controls the wind turbine to idle at a preset low speed (preferably 10%~20% of the rated speed in this embodiment) for a preset time (preferably 5~10 minutes in this embodiment), causing the ice layer attached to the blade surface to fall off through the mechanical vibration generated by the blade rotation and airflow friction. If the wind turbine is equipped with a blade heating device, the blade heating is simultaneously activated during the pre-start procedure.

[0044] When the actual power generation of the wind turbine is detected to be continuously lower than a preset percentage (preferably 70% in this embodiment) of the theoretical expected power based on the current wind speed, it is determined that the blades are icy or covered with snow, and the EMS automatically triggers a de-icing procedure. The de-icing procedure includes increasing the wind turbine speed to 80%~90% of the rated speed and running it continuously for a preset time, or starting the blade heating device until the power generation recovers to more than 90% of the theoretical expected power.

[0045] The electrical energy consumed by the anti-icing pre-start procedure and the de-icing procedure is included in the wind turbine operating energy consumption in the hourly electricity load data. Through the above-mentioned anti-icing and de-icing mechanisms, the wind turbines can reliably start and operate efficiently in high-altitude and low-temperature environments, ensuring the effective output of the wind power generation system under winter icing conditions, and avoiding the impact on the accuracy of the estimated remaining flight time and the staged discharge planning caused by blade icing, which would lead to lower-than-expected wind power output.

[0046] For hourly photovoltaic power generation data, attenuation corrections are made based on the degree of dust accumulation and snow cover on the surface of the photovoltaic panels.

[0047] For hourly electricity load data, when the altitude is higher than 2500m, ventilation compensation load correction is performed, which includes the following sub-steps: S21. Based on the altitude of the current deployment point, query the preset altitude-minimum safe air exchange rate mapping table to determine the target air exchange rate corresponding to the current deployment point. The altitude-minimum safe air exchange rate mapping table is calibrated based on the minimum ventilation required to maintain the oxygen concentration inside the public toilet at different altitudes above a preset safe concentration threshold (19.5% in this embodiment). The target air exchange rate increases non-linearly with increasing altitude.

[0048] S22, the difference between the target number of air changes and the preset benchmark number of air changes (10 times / hour at standard altitude) is multiplied by the unit air change power of the public toilet exhaust system (0.5kW / time) to obtain the ventilation compensation load increment; wherein, the unit air change power is obtained by correcting the rated unit air change power under standard air density according to the ratio of the actual air density at the current deployment point to the standard air density, so as to reflect the change in shaft power required for the fan to deliver the same volume flow rate under high altitude and low air pressure conditions.

[0049] S23, the ventilation compensation load increment is superimposed on each time period of the hourly electricity load data to obtain the corrected hourly electricity load data, and the ventilation compensation load increment is marked as a non-reducible safe load. The non-reducible safe load mark is used to reserve it as the highest priority in the graded discharge plan.

[0050] The corrected three types of data are used as input to the prediction model, so that the subsequent prediction results are not affected by the differences in the deployment site environment.

[0051] S3, input the corrected corresponding data, SOC data, energy storage battery temperature data, meteorological data and scene feature vector into the pre-trained VMD-BiLSTM multi-branch collaborative prediction model, and output the prediction results, specifically including hourly electricity load prediction, hourly photovoltaic power generation prediction, hourly wind power generation prediction, SOC trajectory prediction and energy storage battery temperature prediction for the next 24 hours; the VMD-BiLSTM multi-branch collaborative prediction model includes a VMD decomposition layer, a shared BiLSTM encoder and a multi-branch decoder.

[0052] The SOC trajectory prediction value includes the predicted trend of SOC changes in the energy storage battery at various times within the next 24 hours. Based on this predicted trajectory, the system can predict in advance when the energy storage SOC will drop to a critical value under the current power load and wind and solar power generation conditions, thereby proactively formulating a scheduling strategy before the power is exhausted. Adapting to this scenario, in waterless public toilets, the biodegradation treatment unit needs continuous heating to maintain microbial activity. If the heating device unexpectedly loses power due to SOC prediction deviation, the degradation process will halt, and even after power is restored, it will take several hours to re-establish stable processing capacity. This application predicts the energy storage power decline trend in advance through SOC trajectory prediction, enabling the system to proactively adjust the scheduling strategy before the SOC drops to a critical value, ensuring continuous power supply to critical loads such as biodegradation heating, and avoiding interruptions in the degradation process due to inaccurate predictions or scheduling delays. This solves the technical challenge of continuous heating for stable biodegradation in off-grid operation of waterless public toilets.

[0053] Furthermore, Figure 2This is a flowchart of the training steps for the VMD-BiLSTM multi-branch collaborative prediction model in this embodiment. Before the model is used, it also includes the model training process, such as... Figure 2 As shown, the specific steps of training the VMD-BiLSTM multi-branch collaborative prediction model include: S01, Obtain the training dataset, which includes hourly electricity load data, hourly photovoltaic power generation data, hourly wind power generation data, SOC data, energy storage battery temperature data, meteorological data, scene feature vectors, and the true value labels corresponding to each prediction branch.

[0054] In a preferred embodiment, the scene feature vector further includes an adaptive update mechanism for operating status. During the operation of the target public toilet, the EMS monitors the deviation between the actual power load and the predicted load in real time. When the deviation continuously exceeds a preset threshold, the scene feature vector is automatically triggered for verification: the system re-identifies the regional and seasonal types of the current power consumption mode based on the actual operating data of the most recent preset number of days (e.g., 7 days) and automatically updates the corresponding scene labels; if the identification result is inconsistent with the current setting, a prompt message is generated for maintenance personnel to confirm and update. In addition, the scene feature vector also includes event labels to identify "ordinary days," "holidays," or "large-scale event days" to adapt to the characteristics of peak instantaneous visitor flow in scenic spots. After receiving event information through a preset calendar or remote management platform, the EMS automatically updates the event labels. When the prediction model receives the "holiday" or "large-scale event day" label, it applies a preset load amplification factor to the power load prediction branch.

[0055] S02, construct the initial VMD-BiLSTM multi-branch collaborative prediction model, including VMD decomposition layer, shared BiLSTM encoder and multi-branch decoder.

[0056] The VMD decomposition layer decomposes the input multi-source time-series data into multiple intrinsic mode function (IMF) components. Specifically, for hourly electricity load data, hourly photovoltaic power generation data, and hourly wind power generation data from historical periods, the VMD decomposition layer decomposes each type of data into K IMF components, each IMF component representing the oscillation mode of the original signal at different frequency scales. Through VMD decomposition, the non-stationary original time-series data is transformed into multiple relatively stationary mode components, reducing the learning difficulty of subsequent time-series prediction models.

[0057] In a preferred embodiment, after the VMD decomposition layer decomposes each type of data into multiple IMF components, these components can be classified into three categories—high-frequency components, mid-frequency components, and low-frequency components—based on their frequency characteristics. High-frequency components correspond to rapid random fluctuations in load data (such as instantaneous changes caused by equipment start-up and shutdown), mid-frequency components correspond to intraday periodic changes (such as morning and evening peak hours), and low-frequency components correspond to long-term trend changes (such as seasonal electricity consumption patterns). Different frequency components are input into a shared BiLSTM encoder for feature extraction using a matching prediction time step, avoiding the limitation of a single time step failing to capture multi-scale features simultaneously.

[0058] A shared BiLSTM encoder is used to extract temporal features from each IMF component. The shared BiLSTM encoder employs a bidirectional long short-term memory network structure to perform temporal modeling of each IMF component from both forward and backward directions, capturing the dependencies between components at different time scales. The output of the shared BiLSTM encoder is a shared feature vector that integrates multi-scale temporal information.

[0059] A multi-branch decoder is used to map the shared feature vector to the prediction results of each physical quantity. The multi-branch decoder includes an electricity load prediction branch, a photovoltaic power generation prediction branch, a wind power generation prediction branch, a State of Charge (SOC) trajectory prediction branch, and an energy storage battery temperature prediction branch. Each branch outputs hourly predicted values ​​for the next 24 hours of electricity load, photovoltaic power generation, wind power generation, SOC trajectory, and energy storage battery temperature, respectively. Each decoding branch is an independent fully connected network, responsible for adapting the shared feature vector to the output space of its respective prediction target.

[0060] Through the collaborative work of the VMD decomposition layer, shared BiLSTM encoder, and multi-branch decoder, this model can share underlying temporal features across five prediction tasks while retaining the independently adapted output layer parameters for each task, achieving multi-task joint prediction. The functions of the VMD decomposition layer, shared BiLSTM encoder, and multi-branch decoder are identical during both the training and usage phases. During training, model parameters for each layer are obtained through iterative training from S03 to S05. During usage, the trained parameters are directly loaded to predict the input data, eliminating the need to rebuild the model.

[0061] In a preferred embodiment, considering that the number of modal layers K and the penalty factor α of the VMD decomposition layer are hyperparameters, their values ​​directly affect the decomposition effect and thus the subsequent prediction accuracy. Furthermore, the spectral characteristics of load, photovoltaic, and wind power data differ at different deployment sites, and fixed default values ​​cannot adapt to all scenarios. Therefore, the number of modal layers and the penalty factor of the VMD decomposition layer in step S02 are determined through the following pre-calibration sub-step: Power spectral density estimation was performed on the historical hourly electricity load data, hourly photovoltaic power generation data and hourly wind power generation data in the training dataset, and the number of significant peaks and the bandwidth of each significant peak were extracted for each data point. The modal layer number search space is determined based on the number of significant peaks, and the penalty factor search space is determined based on the bandwidth of each significant peak. Within the combined search domain formed by the modal layer number search space and the penalty factor search space, the optimal combination of modal layer number and penalty factor is searched using a Bayesian optimization algorithm with the objective of minimizing the VMD decomposition and reconstruction error, and then configured in the VMD decomposition layer. The VMD decomposition and reconstruction error is the mean square error between the signal reconstructed from all IMF components after VMD decomposition using candidate parameter combinations and the original signal.

[0062] The aforementioned pre-calibration process enables the VMD decomposition layer to adaptively determine the optimal parameter combination based on the frequency distribution characteristics of the three types of data at different deployment points, avoiding decomposition mismatch caused by manual experience settings, thereby improving the subsequent prediction accuracy.

[0063] S03, the training dataset is processed sequentially by the VMD decomposition layer and the shared BiLSTM encoder, and then input into the multi-branch decoder for prediction to obtain the prediction values ​​of each branch.

[0064] S04, Total Loss Function. Considering the varying severity of consequences of prediction errors in off-grid scheduling for different prediction branches, the deviation in electricity load prediction directly affects the accuracy of remaining flight time estimation, and the deviation in SOC trajectory prediction directly leads to inaccurate triggering timing for graded discharge planning. The safety costs of these two branches in off-grid scenarios are significantly higher than other branches. Conventional multi-task learning typically treats all branches with equal weight, making it difficult to reflect these differences. Therefore, in a preferred embodiment, a total loss function is constructed that includes branch prediction loss terms and additional constraint terms.

[0065] In this embodiment, the branch prediction loss term is a weighted sum of the loss functions of each branch, and the initial value of the loss weight of each branch is determined according to the preset safety cost level of the corresponding branch. The safety cost level is used to characterize the degree of impact of the prediction deviation of the corresponding branch on the reliability of the power supply and equipment safety of the waterless public toilet. In this embodiment, the safety cost level of the electricity load prediction branch and the SOC trajectory prediction branch is higher than that of the photovoltaic power generation prediction branch, the wind power generation prediction branch, and the energy storage battery temperature prediction branch. In practical applications, the above safety cost level can be further divided into more levels according to specific scenarios, and is not limited to the above two levels.

[0066] During training, every preset training epoch (preferably every 5 training epochs in this embodiment), the loss weights are dynamically adjusted based on the rate of change of prediction error for each branch in the most recent validation period. The dynamic adjustment range of the loss weights is constrained by a lower limit, which is determined by the minimum guaranteed weight corresponding to the safety cost level of each branch. That is, the weights of the load prediction branch and the SOC trajectory prediction branch must not be lower than the preset minimum guaranteed weight throughout the entire training process, ensuring that the model always prioritizes the prediction accuracy of these two types of high safety cost branches.

[0067] Furthermore, considering that a purely data-driven prediction model may output physically unreasonable prediction results under extreme conditions of insufficient training data coverage (e.g., the SOC change exceeds the physical upper limit allowed by the maximum charge / discharge rate of the battery, or the photovoltaic output exceeds the theoretical maximum value determined by solar resources), in this embodiment, the additional constraint term includes a physical constraint regularization term, which is used to encode the physical operating laws of the energy storage battery and photovoltaic module in the waterless public toilet microgrid into the training process in the form of a differentiable penalty term, specifically including: The SOC dynamics constraint regularization term is used to impose a dynamic consistency constraint on the SOC trajectory prediction output of the SOC trajectory prediction branch. A penalty value is generated when the SOC change in adjacent time periods in the predicted SOC trajectory exceeds the upper limit of SOC change determined by the maximum charge and discharge rate of the energy storage battery.

[0068] The photovoltaic power generation upper bound constraint regularization term is used to constrain the physical upper bound of the hourly photovoltaic power generation prediction value output by the photovoltaic power generation prediction branch, and generates a penalty value when the hourly photovoltaic power generation prediction value exceeds the theoretical maximum photovoltaic power generation calculated based on the light intensity in meteorological data.

[0069] The energy conservation constraint regularization term is used to impose physical consistency constraints on the energy conversion process of the energy storage battery. A penalty value is generated when the SOC change and charging / discharging power in adjacent time periods do not satisfy a preset energy conversion efficiency relationship for the energy storage battery in the SOC trajectory prediction. The charging / discharging power is calculated based on hourly predicted electricity load, hourly predicted photovoltaic power generation, and hourly predicted wind power generation.

[0070] In this embodiment, the physical constraint regularization term is added to the total loss function as a soft constraint. Its regularization coefficient is set to a first value (preferably 0.1 in this embodiment) in the early stages of training, and gradually decreased to a second value (preferably 0.001 in this embodiment) according to a preset decay strategy as the training epochs increase. The first value is greater than the second value. The purpose of the decay strategy is to guide the model to learn a solution space that satisfies physical laws with strong physical constraints in the early stages of training, and to gradually relax the constraints in the later stages to allow the model to finely fit in the data-driven direction.

[0071] The above-mentioned method of constructing the total loss function enables the model to allocate learning resources differently according to the safety cost level of each branch during training, and to ensure that the prediction results always meet the physical operation law of energy storage batteries and photovoltaic modules through physical constraint regularization terms. This effectively solves the problem that pure data-driven models may output physically infeasible predictions under extreme off-grid conditions, and improves the prediction credibility and safety of the model under conditions with few samples and across scenarios.

[0072] S05, with the goal of minimizing the total loss function, the parameters of the initial VMD-BiLSTM multi-branch collaborative prediction model are iteratively optimized using a preset training strategy until the preset convergence condition is met, thus obtaining the trained VMD-BiLSTM multi-branch collaborative prediction model.

[0073] Considering that portable toilets exhibit both commonalities (such as similar basic load patterns for lighting, ventilation, and biodegradation treatment) and significant differences (such as differences in underfloor heating / air conditioning loads between frigid and non-frigid regions, and the impact of different altitudes on ventilation loads) when using the same learning rate for both the encoder and decoder, two problems may arise: an excessively high learning rate may destroy the cross-scene common features already learned by the encoder, while an excessively low learning rate may prevent the decoder from quickly adapting to the unique features of new scenarios. Therefore, in this embodiment, during parameter iterative optimization, the parameters of the shared BiLSTM encoder use a first learning rate, while the parameters of each branch of the multi-branch decoder use a second learning rate, with the first learning rate being lower than the second learning rate. This differentiated learning rate setting ensures that the encoder does not lose the cross-scene common temporal features learned during the pre-training phase when adapting to new scenarios, while the decoder can quickly adapt to the current scenario's power consumption patterns with a small amount of new data.

[0074] Furthermore, considering the mobile nature of waterless public toilets, the same set of equipment may be relocated to scenic areas with different altitudes and climates in different seasons. After each relocation, the historical data accumulation at the new deployment point is almost zero. If the model needs to be trained from scratch after each relocation, the time and data costs would be too high. In this embodiment, the preset training strategy specifically includes a pre-training stage and a fine-tuning stage: During the pre-training phase, the basic parameters of the shared BiLSTM encoder and each prediction branch are trained based on the training dataset (which includes historical operation data of multiple deployed public toilets), enabling the VMD-BiLSTM multi-branch collaborative prediction model to learn the cross-scenario common time-series characteristics of the public toilet wind-solar-storage microgrid.

[0075] During the fine-tuning phase, when the target public toilet is relocated to the new deployment site, the following sub-steps are performed: S051, freeze the high-level parameters of the shared BiLSTM encoder, and only fine-tune the parameters of each branch in the multi-branch decoder and the low-level parameters of the shared BiLSTM encoder. The training data used for fine-tuning is the operational data collected at the newly deployed location of the target public toilet. The reason for freezing the high-level parameters here is that the high-level parameters have encoded general temporal patterns across scenarios. If the amount of data in the new scenario is limited, and the high-level parameters are updated significantly at the same time, it may cause the model to forget the general knowledge learned in the pre-training stage, resulting in overfitting.

[0076] S052, when the cumulative operating data duration of the target public toilet at the newly deployed site reaches a preset data duration threshold (preferably 7 days in this embodiment), the parameter freeze is lifted and all parameters of the shared BiLSTM encoder and the multi-branch decoder are jointly fine-tuned. At this time, the amount of data at the newly deployed site is sufficient to support the model in finely adjusting all parameters, enabling the model to further adapt to the personalized electricity consumption characteristics of the current deployment site while maintaining its cross-scenario generality.

[0077] The two-stage transfer training strategy described above enables the model to learn common features across scenarios during the pre-training stage, and then quickly adapt to the unique features of the new scenario through fine-tuning after the model is transferred to a new deployment point. As data accumulates, the model gradually transitions from partial fine-tuning to full parameter fine-tuning, balancing the need for rapid deployment in the early stages of the transfer with the need for accuracy in long-term operation. This effectively solves the problems of difficulty in accumulating historical data and decreased prediction accuracy after model transfer caused by frequent changes in the deployment points of mobile toilets.

[0078] In another preferred embodiment, to further shorten the model adaptation cycle at new deployment sites, a meta-learning framework is used instead of conventional pre-training in the pre-training phase. Specifically, the initial parameters of the model are trained using a meta-learning algorithm (such as MAML) on training tasks at multiple historical deployment sites, enabling the model to acquire meta-knowledge to quickly adapt to new scenarios. When the target public toilet is moved to a new deployment site, only 1-3 days of operational data need to be collected, and a small amount of gradient updates can be quickly fine-tuned based on the initial parameters. It can be put into use without waiting for data accumulation to reach 7 days, further improving the flexibility and timeliness of mobile public toilet deployment.

[0079] S4. Determine the remaining energy capacity based on the prediction results and the pre-acquired weather forecast data, and formulate a graded discharge plan for the energy storage system of the target public toilet based on the SOC data and the remaining energy capacity. The graded discharge plan is divided into multiple energy capacity levels according to the remaining energy capacity, and each energy capacity level corresponds to a preset load reduction ratio.

[0080] Considering that sunlight intensity and wind speed have the greatest impact on photovoltaic and wind power output in weather forecasts, with sunlight intensity directly determining the upper limit of photovoltaic power generation and wind speed directly determining the output power of wind turbines, in this embodiment, the weather forecast data includes at least daily sunlight intensity forecasts and hourly wind speed forecasts for the next 24 hours.

[0081] Specifically, determining the number of subsequent flights based on the prediction results and the pre-acquired weather forecast data includes: Based on the forecast results, calculate the net energy value (forecasted photovoltaic output + forecasted wind power output - forecasted load) for each period in the next 24 hours, and determine whether there is a net energy deficit (i.e., periods with negative net energy values). When a net energy deficit exists, divide the difference between the current SOC data and the pre-acquired lower limit of SOC by the daily average value of the net energy deficit to obtain an initial estimate of the remaining space capacity.

[0082] In a preferred embodiment, the VMD-BiLSTM multi-branch collaborative prediction model outputs a confidence interval corresponding to the predicted SOC trajectory value, along with the predicted value. This confidence interval is obtained using an interval prediction method and is used to quantify the uncertainty of the prediction result. The wider the interval, the lower the model's confidence in the predicted SOC value at that moment. During the model training phase, the confidence interval is output synchronously as part of the prediction results of each branch in step S03. In subsequent steps S04, it serves as the basis for calculating the total loss function and for parameter iteration optimization in step S05. During the usage phase, the confidence interval serves as the basis for secondary reduction in step S4. This preferred embodiment does not conflict with the aforementioned conventional training method and represents a further refinement within the same model architecture.

[0083] Furthermore, based on the forecasts of consecutive days of low radiation (light intensity consistently below a preset threshold) and / or consecutive days of low wind speed (wind speed consistently below the wind turbine cut-in wind speed), a meteorological risk reduction is applied to the initial estimate of the remaining spaceflight duration. The longer the forecast of consecutive days of severe weather, the greater the reduction. Simultaneously, a second reduction is applied to the meteorological reduction result based on the width of the uncertainty interval in the SOC trajectory prediction. A wider uncertainty interval indicates lower prediction reliability, a smaller reduction coefficient, and a more conservative scheduling decision, ultimately yielding the remaining spaceflight duration.

[0084] After obtaining the remaining battery life, a tiered discharge plan is formulated based on the SOC data and the remaining battery life. Considering the impact of battery aging on the actual usable capacity, as the number of cycles increases, the battery SOH decreases, and the actual usable energy corresponding to the same SOC value decreases. In this embodiment, the capacity decay ratio is determined based on the SOH data, and the SOC trajectory prediction is corrected by capacity conversion to obtain a capacity-corrected SOC trajectory. Based on the lowest predicted point of the capacity-corrected SOC trajectory in the next 24 hours and the remaining battery life, multiple battery life levels are generated (e.g., ≥3 days for Level 1, 1~3 days for Level 2, <1 day for Level 3).

[0085] In a preferred embodiment, the trigger condition for the graded discharge plan is as follows: when the remaining power supply is lower than a preset safety threshold (preferably 1.5 days in this embodiment), the graded discharge plan is automatically initiated. The multiple power supply levels are divided into four levels according to the importance of the load: Level 1 is the full-load operation level, where all electrical equipment is powered normally; Level 2 is the non-essential load shutdown level, cutting off non-core equipment such as landscape lighting and information display screens; Level 3 is the anti-freeze and ventilation level, retaining only biodegradable heating, exhaust systems, and basic lighting; Level 4 is the emergency insulation level, retaining only the biodegradable heating load and the non-reducible safety load marked in step S2 (the incremental portion of the ventilation compensation load), while all other loads are cut off. The load reduction ratio corresponding to each level increases progressively with the shortening of the remaining power supply; the shorter the remaining power supply, the higher the execution level and the less load is retained.

[0086] In a preferred embodiment, when meteorological data detects that the decrease in light intensity exceeds a preset abrupt change threshold within a preset time window (e.g., within one hour), the system triggers an extreme weather emergency mode: skipping the daily assessment process for the remaining flight time as usual, it directly calculates the hourly remaining flight time based on the real-time ratio of the current State of Charge (SOC) to the net energy deficit, and immediately upgrades the scheduling level by one level for execution, without waiting for the next scheduling cycle. Simultaneously, during the emergency mode, scheduling decisions are primarily based on the current measured net energy value, with prediction results only serving as a reference for subsequent trends, avoiding delays in response due to inaccurate prediction models under extreme conditions.

[0087] When the SOC data decreases sequentially to the trigger threshold corresponding to each range level, a preset load reduction is performed according to the corresponding range level until only the non-reducible safe load remains. This refers to the non-reducible safe load (the incremental portion of the ventilation compensation load) marked in step S2.

[0088] In another preferred embodiment, considering that waterless public toilets are mostly deployed in remote areas with unstable communication conditions, the BMS, weather stations, or smart meters may experience communication interruptions or abnormal data collection. If the system completely stops scheduling decisions due to partial data loss, the power supply to the public toilets will be out of control. In this embodiment, the EMS has a built-in data integrity monitoring module. When it detects that the collected values ​​of any key data source (including SOC data, photovoltaic output data, wind power output data, meteorological data, and load measurement data) are continuously missing for more than a preset time (preferably 15 minutes in this embodiment) or the values ​​exceed the physically reasonable range, the following graded degradation strategy is executed: (1) If meteorological data is missing, the most recent valid meteorological data is used as the static input, and the meteorological risk reduction factor of the continued spaceflight number is set to the maximum value to make the continued spaceflight number estimate more conservative; (2) If SOC data is missing, the estimated consumption based on load forecast during the missing period is subtracted from the most recent valid SOC value as the replacement value, and the current scheduling result is marked as "low confidence". The scheduling level is automatically raised by one level for execution. (3) If photovoltaic or wind power output data is missing, a physical mechanism model based on current measured meteorological data shall be used to replace the measured values. The physical mechanism model specifically includes: Photovoltaic power generation = Rated power of photovoltaic module × (Measured light intensity / Irradiance under standard test conditions) × Temperature correction factor × Dust attenuation factor; Wind power generation capacity = Table lookup value of the standard power curve of the wind turbine at the current measured wind speed × (measured air density / standard air density); The dust attenuation coefficient and temperature correction coefficient are taken from the most recent valid correction parameters in step S2. An additional uncertainty margin (preferably 20% of the theoretical output) is applied to the net energy deficit calculation. If the current measured meteorological data is synchronously missing, the most recent valid meteorological data is used to replace the measured data.

[0089] (4) If the measured load data is missing, the load forecast value output in step S3 is used instead, and the uncertainty range of the load forecast is expanded to a preset multiple in the graded discharge planning. The above degradation strategy ensures that the system can still maintain basic scheduling capabilities when some sensors fail or communication is interrupted, avoiding complete scheduling paralysis due to a single data source failure, and improving the robustness of the system in harsh deployment environments. When the missing data resumes normal collection and continues for more than the preset verification time (preferably 30 minutes in this embodiment), the system automatically exits the degradation mode and resumes the normal scheduling process.

[0090] Furthermore, considering the electrochemical characteristics of lithium-ion batteries at low temperatures, increased electrolyte viscosity and decreased lithium-ion diffusion rate lead to a significant increase in battery internal resistance and a decrease in usable capacity. If the battery temperature drops below 0°C, its discharge capacity may only be 60% to 80% of that at room temperature; when the temperature drops below -20°C, the usable capacity may even be less than 50% of that at room temperature. In such low-temperature scenarios, the energy storage battery not only needs to supply power externally but also needs to consume its own energy for electric heating to raise the cell temperature to a suitable operating range (usually 15°C to 35°C). If sufficient energy is not reserved, the battery may experience a rapid drop in SOC to a critical value due to self-heating energy consumption at low temperatures, thereby triggering the system undervoltage protection and forcibly cutting off power, completely interrupting the power supply to the public toilet.

[0091] In this embodiment, in the graded discharge planning, the lower limit of SOC of the energy storage battery is dynamically adjusted according to the current ambient temperature: when the ambient temperature is lower than the preset low temperature threshold (preferably -10℃ in this embodiment), the lower limit of SOC is adjusted from the normal value (20% in this embodiment) to the preset low temperature protection value (preferably 50% in this embodiment) so as to reserve the energy required for self-heating in the energy storage battery in the low temperature environment.

[0092] When the ambient temperature is higher than or equal to the preset low temperature threshold (i.e., ≥-10℃), the SOC lower limit returns to the normal value (20%), the graded discharge plan is executed in normal mode, and the trigger thresholds of each range level are restored accordingly, without reserving additional self-heating energy space. This temperature hysteresis mechanism can effectively avoid frequent switching of the SOC lower limit near the low temperature threshold, ensuring the stability of the scheduling strategy.

[0093] When the ambient temperature rises above the preset low temperature threshold, the SOC lower limit automatically returns to the normal value.

[0094] The aforementioned tiered discharge planning integrates multi-dimensional information such as SOC data, weather forecasts, SOC trajectory prediction uncertainties, SOH decay, and temperature to achieve prediction under off-grid conditions. Instead of passively cutting off loads when the stored energy is about to be exhausted, the system predicts in advance the number of days it can maintain operation based on forecast information and proactively divides the system into multiple endurance levels with preset load reduction ratios for each level. This makes the discharge process more refined and orderly, effectively avoiding power outages and shutdowns of public toilets due to premature depletion of stored energy, and ensuring the power supply reliability and continuous operation capability of waterless public toilets under off-grid conditions.

[0095] In a preferred embodiment, considering that lithium batteries need to consume energy for self-heating in low-temperature environments to maintain normal operating temperature, and that the electrical energy consumed by heating itself directly affects the available capacity and range estimation of the energy storage system, this embodiment incorporates the battery heating power as a scheduling decision variable into the optimization model of the graded discharge planning.

[0096] Specifically, when the ambient temperature is lower than a preset low-temperature threshold (preferably -10℃ in this embodiment), the graded discharge planning simultaneously calculates the minimum and recommended power required for battery self-heating at the current temperature when determining the load reduction ratio for each range level. The minimum heating power is the minimum electrical power required to maintain the battery temperature above the minimum operating temperature; the recommended heating power is the electrical power required to raise the battery temperature to the optimal operating temperature range within a preset time.

[0097] The scheduling of heating power follows the following priority principle: when the energy storage SOC is higher than the preset minimum protection threshold (preferably 30% in this embodiment), the heating power has the same power supply priority as the basic load of the public toilet (lighting, ventilation, biodegradable heating, etc.) and is included in the load reduction ratio calculation for each range level; when the energy storage SOC is lower than the preset minimum protection threshold, heating is prohibited from starting or the heating power is forcibly reduced to the minimum maintenance level to avoid over-discharge of energy storage due to heating consumption.

[0098] The aforementioned heating power scheduling mechanism works in conjunction with the dynamic adjustment of the lower limit temperature of SOC. The increase of the lower limit of SOC reserves the energy space required for self-heating in low-temperature environments, while the heating power scheduling performs fine-grained control over the heating process within this energy space. The two mechanisms respectively address the coordination problem between battery self-heating and power supply guarantee in low-temperature environments from the perspectives of "how much to reserve" and "how to consume," thus avoiding the heating energy consumption from crowding out the power supply of critical loads or causing premature depletion of energy storage.

[0099] In another alternative implementation, the target public toilet may have external power grid access. In this scenario, the system enters the grid-connected economic dispatch mode: aiming to minimize the total daily operating cost of the system, a day-ahead optimization dispatch model is constructed based on the prediction results output from step S3, solving for the energy storage charging and discharging plans and grid interaction plans for each time period in the next 24 hours. The above mode, together with the off-grid graded discharge planning mode described in step S4, constitutes the dual-mode dispatch framework of this method, making it applicable to both grid-connected and grid-free deployment scenarios.

[0100] In summary, the multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets provided in this embodiment is based on the construction of a complete technical chain from environmental perception, physical correction, multi-task prediction to graded discharge planning. In step S2, ventilation compensation correction is applied to the load based on altitude, photovoltaic output is corrected based on dust accumulation and snow cover, and wind power output is corrected based on air density. These three types of corrections serve as a unified pre-processing layer, eliminating the impact of environmental differences on the prediction model input from the data source. This allows the model to obtain accurate prediction results without retraining when migrating between different deployment points. The VMD-BiLSTM multi-branch collaborative prediction model uses a two-stage training strategy of pre-training and fine-tuning. In the pre-training stage, it learns common features across scenarios, and in the fine-tuning stage, it quickly adapts to the individual characteristics of new deployment points through differentiated learning rates and progressive unfreezing strategies. This solves the problem of frequent changes in the deployment points of mobile public toilets. The model addresses the challenges of accumulating historical data and declining prediction accuracy after model migration. Simultaneously, by employing a loss function weighted by safety costs and physical constraint regularization, the model prioritizes the accuracy of high-risk prediction branches during training while ensuring that prediction results consistently conform to the physical operating laws of energy storage batteries and photovoltaic modules, thus enhancing the model's prediction reliability under extreme off-grid conditions. Finally, by fusing prediction results, weather forecast data, SOC trajectory prediction uncertainty, and SOH decay information, the model uses meteorological risk reduction and uncertainty secondary reduction to determine the remaining battery life, generating multiple battery life levels and corresponding preset load reduction ratios for each level. This effectively prevents public toilets from shutting down due to premature energy storage depletion.

[0101] The synergistic effect of the above steps enables this method to effectively address challenges such as extreme environmental parameter interference, scene migration due to mobile deployment, and difficulty in ensuring power supply reliability under off-grid conditions in complex environments like plateaus, Gobi Desert, and arid lands. It maximizes off-grid operation time while ensuring the basic sanitation functions of public toilets. Compared to the crude control schemes of existing technologies that simply combine photovoltaic panels, wind turbines, and batteries, this method achieves a leap from passive response to forward prediction. Furthermore, through the deep integration of physical constraints and data-driven approaches, it enhances the model's prediction reliability and generalization ability under conditions of limited samples and cross-scenario application.

[0102] Example 2: Finally, this application also proposes a computing device, which includes a processor and a memory. The memory stores computer programs, and the processor executes the instructions stored in the memory so that the computer device can execute the multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets described in the above embodiments.

[0103] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets, characterized in that, include: S1, acquire multi-source operation data, scene feature vector and environmental parameters of the target public toilet. The multi-source operation data includes hourly electricity load data, hourly photovoltaic power generation data, hourly wind power generation data, SOC data, energy storage battery temperature data and meteorological data. S2, the hourly electricity load data, hourly photovoltaic power generation data and hourly wind power generation data are corrected according to the environmental parameters to obtain the corrected corresponding data; S3, input the corrected corresponding data, SOC data, energy storage battery temperature data, meteorological data, and scene feature vector into the pre-trained VMD-BiLSTM multi-branch collaborative prediction model, and output the prediction results. The prediction results include hourly electricity load predictions, hourly photovoltaic power generation predictions, hourly wind power generation predictions, SOC trajectory predictions, and energy storage battery temperature predictions for the next 24 hours. The VMD-BiLSTM multi-branch collaborative prediction model includes a VMD decomposition layer, a shared BiLSTM encoder, and a multi-branch decoder. S4. Determine the remaining power capacity based on the prediction results and the pre-acquired weather forecast data, and formulate a graded discharge plan based on the SOC data and the remaining power capacity. The graded discharge plan divides the remaining power capacity into multiple power capacity levels, and each power capacity level corresponds to a preset load reduction ratio.

2. The multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets according to claim 1, characterized in that, The training process of the VMD-BiLSTM multi-branch collaborative prediction model includes: S01, Obtain the training dataset, which includes hourly electricity load data, hourly photovoltaic power generation data, hourly wind power generation data, SOC data, energy storage battery temperature data, meteorological data, scene feature vectors, and the true value labels corresponding to each prediction branch. S02, Construct an initial VMD-BiLSTM multi-branch collaborative prediction model, which includes a VMD decomposition layer, a shared BiLSTM encoder, and a multi-branch decoder; the multi-branch decoder includes an electricity load prediction branch, a photovoltaic power generation prediction branch, a wind power generation prediction branch, a SOC trajectory prediction branch, and an energy storage battery temperature prediction branch. S03, the training dataset is processed sequentially by the VMD decomposition layer and the shared BiLSTM encoder, and then input into the multi-branch decoder for prediction to obtain the prediction values ​​of each branch; S04, Construct the total loss function that includes the branch prediction loss term and the additional constraint term; S05, with the goal of minimizing the total loss function, the parameters of the initial VMD-BiLSTM multi-branch collaborative prediction model are iteratively optimized using a preset training strategy until the preset convergence condition is met, thus obtaining the trained VMD-BiLSTM multi-branch collaborative prediction model; wherein, during the parameter iterative optimization process, the parameters of the shared BiLSTM encoder adopt a first learning rate, and the parameters of each branch of the multi-branch decoder adopt a second learning rate, wherein the first learning rate is lower than the second learning rate.

3. The multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets according to claim 2, characterized in that, The branch prediction loss term is the weighted sum of the loss functions of each branch. The initial value of the loss weight of each branch is determined according to the preset safety cost level of the corresponding branch. The safety cost level is used to characterize the degree of influence of the prediction deviation of the corresponding branch on the reliability of the power supply and equipment safety of the waterless public toilet. Among them, the safety cost level of the electricity load prediction branch and the SOC trajectory prediction branch is higher than that of the photovoltaic power generation prediction branch, the wind power generation prediction branch and the energy storage battery temperature prediction branch. During training, every preset training round, the loss weight is dynamically adjusted based on the rate of change of prediction error of each branch in the most recent verification period; the dynamic adjustment range of the loss weight is constrained by a lower limit, which is determined by the minimum guarantee weight corresponding to the security cost level of each branch.

4. The multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets according to claim 3, characterized in that, The additional constraint terms include physical constraint regularization terms, which specifically include: The SOC dynamic constraint regularization term is used to apply a dynamic consistency constraint to the SOC trajectory prediction output by the SOC trajectory prediction branch. A penalty value is generated when the SOC change in adjacent time periods in the predicted SOC trajectory exceeds the upper limit of SOC change determined by the maximum charge and discharge rate of the energy storage battery. The photovoltaic power generation upper bound constraint regularization term is used to constrain the physical upper bound of the hourly photovoltaic power generation prediction value output by the photovoltaic power generation prediction branch, and generates a penalty value when the hourly photovoltaic power generation prediction value exceeds the theoretical maximum photovoltaic power generation calculated based on the light intensity in the meteorological data. The energy conservation constraint regularization term is used to constrain the physical consistency of energy conversion. When the change in SOC and the charging / discharging power in adjacent time periods do not meet the preset energy conversion efficiency relationship of the energy storage battery in the SOC trajectory prediction, a penalty value is generated. The charging / discharging power is calculated based on the hourly electricity load prediction value, the hourly photovoltaic power generation prediction value, and the hourly wind power generation prediction value.

5. The multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets according to claim 4, characterized in that, In step S05, the preset training strategy specifically includes a pre-training phase and a fine-tuning phase: During the pre-training phase, the basic parameters of the shared BiLSTM encoder and each prediction branch are trained based on the training dataset, enabling the VMD-BiLSTM multi-branch collaborative prediction model to learn the cross-scenario common temporal features of the target public toilet wind-solar-storage microgrid. During the fine-tuning phase, when the target public toilet is relocated to the new deployment site, the following sub-steps are performed: S051, Freeze the high-level parameters of the shared BiLSTM encoder, and only fine-tune the parameters of each branch in the multi-branch decoder and the low-level parameters of the shared BiLSTM encoder. The training data used for fine-tuning training is the operation data of the target public toilet collected at the new deployment point. S052, when the cumulative operating data duration of the target public toilet at the newly deployed site reaches the preset data duration threshold, the parameter freeze is lifted and all parameters of the shared BiLSTM encoder and the multi-branch decoder are jointly fine-tuned.

6. The multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets according to claim 2, characterized in that, The number of modal layers and the penalty factor of the VMD decomposition layer in step S02 are determined through the following pre-calibration sub-step: Power spectral density estimation was performed on the historical hourly electricity load data, hourly photovoltaic power generation data and hourly wind power generation data in the training dataset, and the number of significant peaks and the bandwidth of each significant peak were extracted for each data point. The modal layer number search space is determined based on the number of significant peaks, and the penalty factor search space is determined based on the bandwidth of each significant peak. Within the combined search domain formed by the modal layer number search space and the penalty factor search space, the VMD decomposition and reconstruction error is minimized as the objective, where the VMD decomposition and reconstruction error is the mean square error between the reconstructed signal and the original signal. The optimal combination of modal layer number and optimal penalty factor is searched using a Bayesian optimization algorithm and configured in the VMD decomposition layer.

7. The multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets according to claim 1, characterized in that, In step S2, the step of correcting the hourly electricity load data, hourly photovoltaic power generation data, and hourly wind power generation data according to the environmental parameters to obtain the corrected corresponding data specifically includes: When the altitude in the environmental parameters is higher than a preset high-altitude threshold, ventilation compensation load correction is performed. The ventilation compensation load correction includes the following sub-steps: S21, based on the altitude, query the preset altitude-minimum safe air change rate mapping table to determine the target air change rate corresponding to the current deployment point; wherein, the altitude-minimum safe air change rate mapping table is obtained based on the minimum ventilation volume required to maintain the oxygen concentration inside the public toilet at different altitudes not lower than the preset safe concentration threshold, and the target air change rate increases non-linearly with the altitude. S22, multiply the difference between the target air exchange rate and the preset benchmark air exchange rate by the unit air exchange power of the public toilet exhaust system to obtain the ventilation compensation load increment; wherein, the unit air exchange power is determined according to the actual air density at the current deployment point; S23, the ventilation compensation load increment is superimposed on each time period of the hourly electricity load data to obtain the corrected hourly electricity load data, and the ventilation compensation load increment is marked as a non-reducible safe load. The non-reducible safe load mark is used to reserve it as the highest priority in the graded discharge plan.

8. The multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets according to claim 1, characterized in that, The step of determining the number of subsequent flights based on the prediction results and the pre-acquired weather forecast data specifically includes: The prediction results also include the confidence interval of the SOC trajectory prediction value, and the uncertainty interval width is obtained based on the confidence interval; Based on the prediction results, calculate the net energy value for each period in the next 24 hours and determine whether there is a net energy gap. When there is a net energy gap, divide the difference between the current SOC data and the pre-acquired lower limit of SOC by the daily average value of the net energy gap to obtain the initial estimate of the number of days of continued flight. Based on the forecast of consecutive low-irradiance days and / or consecutive low-wind-speed days in the weather forecast data, the initial estimate of the number of days of continued air travel is reduced by meteorological risk; the meteorological reduction result is further reduced based on the uncertainty interval width of the SOC trajectory prediction to obtain the number of days of continued air travel. The larger the uncertainty interval width, the smaller the reduction coefficient.

9. The multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets according to claim 1 or 7, characterized in that, In step S4, a graded discharge plan is formulated based on the SOC data and the remaining lifespan, specifically including: The capacity decay ratio is determined based on the pre-acquired SOH data. The SOC trajectory prediction is then corrected by capacity conversion based on the capacity decay ratio to obtain the capacity-corrected SOC trajectory. Based on the lowest predicted point of the capacity-corrected SOC trajectory in the next 24 hours and the remaining flight time, multiple flight duration levels are generated. When the SOC data decreases sequentially to the trigger threshold corresponding to each battery life level, a preset load reduction is performed according to the corresponding battery life level until only the highest priority load is retained.

10. The multi-source data prediction and scheduling method for wind-solar-storage microgrids for waterless public toilets according to claim 9, characterized in that, In the graded discharge plan, the preset lower limit of SOC is dynamically adjusted according to the current ambient temperature: When the ambient temperature is lower than the preset low temperature threshold, the SOC lower limit is adjusted to the preset low temperature protection value so as to reserve the energy required for self-heating in the energy storage battery in a low temperature environment.

Images