Electrolytic hydrogen generation buffer system for medical oxygen supply

Abstract

This invention discloses an electrolytic hydrogen production buffer system for medical oxygen supply, specifically relating to the field of electrolytic hydrogen production buffer technology. The system includes acquiring historical data on hospital oxygen consumption, weather forecasts, photovoltaic power generation predictions, and grid electricity prices, as well as current electrolysis operation data, hydrogen storage status data, and power generation operation data. It performs time alignment, cross-source slicing, and feature encoding to output a multi-source state sequence. By constructing future multi-scenario prediction results from hospital oxygen consumption, photovoltaic power supply, electricity prices, and equipment operation data, it collaboratively solves the current hydrogen production, storage, and power generation control strategies based on a multi-scenario common action extraction and rolling update mechanism.

Description

Technical Field

[0001] This invention relates to the field of electrolytic hydrogen production buffer technology, and more specifically, to an electrolytic hydrogen production buffer system for medical oxygen supply. Background Technology

[0002] In the application of electrolytic hydrogen production for medical oxygen supply, the mainstream practice in the industry is to maintain the continuity of oxygen supply while taking into account operating costs under the conditions of fluctuating hospital oxygen demand, photovoltaic power supply fluctuations and electricity price changes. This usually involves collecting historical oxygen consumption data, weather forecast data, photovoltaic power generation forecast data and grid electricity price data, building a prediction model to obtain the oxygen demand and power supply capacity for future periods, and then performing joint scheduling of hydrogen production, hydrogen storage and power generation strategies based on the prediction results. For example, in hospital oxygen supply scenarios equipped with photovoltaic power sources, electrolysis devices, hydrogen storage units, and power generation units, the system needs to simultaneously cope with the gradual oxygen use in general wards, the phased changes in oxygen use caused by surgical scheduling, and the short-term fluctuations in oxygen use caused by ICU and emergency rescue without interrupting the supply. It is also subject to hard constraints such as photovoltaic output being significantly affected by weather changes, time-of-use fluctuations in electricity prices, the inertia of electrolysis device regulation, and the need to maintain a safety margin when calling up hydrogen storage. However, under this constraint, the mainstream approach will consistently reveal an observable and verifiable defect: the system usually generates the current control strategy directly based on a single prediction result or a single path optimization result. When the actual oxygen consumption, actual photovoltaic output, or actual electricity price trend deviates from the prediction, the executed hydrogen production, hydrogen storage, and power generation strategies are prone to mismatch with the actual operating scenario. This manifests as the preceding strategy meeting the oxygen supply continuity and cost constraints under the prediction conditions when it is generated, but after the prediction deviation occurs, it will cause the hydrogen storage to be called up in advance, the electrolysis to be delayed, or the timing of power generation switching to be inaccurate. The reason is that the existing methods lack the ability to extrapolate multiple possible future evolution scenarios in parallel, and also lack the processing mechanism to extract common executable current strategies from multiple future scenarios. The technical problem this application aims to solve is: how to construct future multi-scenario prediction results based on multi-source data in the electrolytic hydrogen production buffer process for medical oxygen supply, and solve the current execution strategy that holds true for multiple scenarios from the future multi-scenario prediction results, so as to improve the adaptability of the coordinated control of hydrogen production, hydrogen storage and power generation to prediction deviations. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an electrolytic hydrogen production buffer system for medical oxygen supply. This system constructs future multi-scenario prediction results based on hospital oxygen consumption, photovoltaic power supply, electricity prices, and equipment operation data. It also collaboratively solves the hydrogen production, hydrogen storage, and power generation control strategies for the current period based on a multi-scenario common action extraction and rolling update mechanism, thereby addressing the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an electrolytic hydrogen production buffer system for medical oxygen supply, comprising a state construction module, a prediction module, a strategy solving module, a cooperative execution module, and an update module. The state construction module is used to acquire historical data on hospital oxygen consumption, weather forecasts, photovoltaic power generation forecasts, and grid electricity prices, as well as current electrolysis operation data, hydrogen storage status data, and power generation operation data. It performs time alignment, cross-source slicing, and feature encoding, and outputs a multi-source state sequence. The prediction module is used to input multi-source state sequences into a temporal attention network, perform joint extraction of oxygen consumption change features, photovoltaic fluctuation features, and electricity price change features, and construct a scenario tree for future periods based on the joint extraction results, outputting oxygen consumption prediction results, power supply prediction results, and price prediction results corresponding to each scenario node; The strategy solving module, based on the oxygen consumption prediction results, power supply prediction results, and price prediction results corresponding to each scenario node, combined with electrolysis operation data, hydrogen storage status data, and power generation operation data, performs a two-level strategy search with the objectives of prioritizing oxygen supply continuity and operating cost constraints, solves the hydrogen production path, hydrogen storage path, and power generation path corresponding to each scenario node, and outputs the scenario strategy set. The collaborative execution module is used to perform scenario backtracking consistency verification and cross-scenario common prefix extraction on the scenario strategy set, generate the current execution strategy, and perform collaborative control on the electrolysis unit, hydrogen storage unit and power generation unit according to the current execution strategy, and output the buffer execution results of the current time period.

[0005] In a preferred embodiment, it further includes: The update module is used to obtain the actual oxygen consumption results, actual photovoltaic results, actual electricity price results, and actual operation results corresponding to the buffer execution results, perform deviation inversion and policy weight update with the scenario policy set, and write the update results back to the temporal attention network and the two-layer policy search, and output the rolling buffer control results for the next time period.

[0006] In a preferred embodiment, the state construction module includes: Perform time mapping and gap filling on hospital oxygen consumption data, weather forecast data, photovoltaic power generation forecast data, power grid price data, electrolysis operation data, hydrogen storage status data, and power generation operation data according to a unified time scale, and output an aligned dataset; Perform adjacent difference calculations on the hospital oxygen value sequence, photovoltaic power generation prediction value sequence and grid electricity value sequence in the aligned dataset, and determine the time when the difference sign changes, the time when the electricity value changes, and the connection time between historical data and current data as slice points. Divide time slices according to the time interval between adjacent slice points and output a multi-source fragment set. For each time slice in the multi-source segment set, perform intra-slice feature extraction, inter-slice difference calculation, and cross-source change association calculation, and concatenate them according to a fixed field order to form a state vector. Arrange the state vectors in chronological order and output the multi-source state sequence.

[0007] In a preferred embodiment, the prediction module includes: The multi-source state sequence is divided into continuous observation windows according to time order and input into a temporal attention network with position encoding and causal mask. The intra-head attention score is calculated for the hospital oxygen consumption field, photovoltaic power generation field and electricity price field in each observation window, and normalized weighted summation is performed and cross-attention fusion is performed with the remaining fields. The output is the oxygen consumption feature sequence, photovoltaic feature sequence, electricity price feature sequence and shared state sequence. Based on the oxygen consumption feature sequence, photovoltaic feature sequence, electricity price feature sequence and shared state sequence, a joint transition tensor for each future time is constructed. For each future time, the oxygen consumption state distribution, power supply state distribution and price state distribution are solved respectively. The three types of state distributions at the same future time are combined into candidate scenario nodes according to the time correspondence relationship. The set of candidate scenario nodes and the joint probability value of each candidate scenario node are output.

[0008] In a preferred embodiment, the prediction module further includes: For each candidate scene node in the candidate scene node set, calculate the node cost value, which consists of the logarithm of conditional probabilities, the sum of squared state transition differences, and the sum of squared field reconstruction differences. Sort the candidate scene nodes under the same parent node according to the node cost value, retain the first node in the sort, and write back the field differences of the remaining candidate scene nodes to the shared state sequence. Repeat the joint transition tensor construction and node cost value calculation on the shared state sequence after writing back, until the first node in the sorting is consistent in two consecutive rounds and the field difference vector is consistent. Output the corrected scene node set. In the order of future time, each correction scenario node in the correction scenario node set is connected to its corresponding parent node in the previous time step. The joint probability value of each connection is multiplied layer by layer, and the multiplication result of the nodes in the same layer is normalized to generate a scenario tree. The field values ​​corresponding to the oxygen consumption state distribution, power supply state distribution and price state distribution in each scenario node are read, and the oxygen consumption prediction result, power supply prediction result and price prediction result corresponding to each scenario node are output.

[0009] In a preferred embodiment, the strategy solving module includes: The oxygen consumption prediction results, power supply prediction results, and price prediction results corresponding to each scenario node are aligned with the electrolysis operation data, hydrogen storage status data, and power generation operation data according to time. A joint state point set consisting of electrolysis power, hydrogen storage flow, and power generation is constructed. The joint state point set is then connected according to the conservation of power change, hydrogen storage balance, and oxygen and power supply balance between adjacent time points. The candidate path graph corresponding to each scenario node is then output. For each path in each candidate path graph, calculate the continuity cost consisting of the sum of squares of the oxygen supply gap, the sum of squares of the hydrogen storage overrun, and the sum of squares of the state jump variables at each time point. Sort and retain each path in ascending order of continuity cost. Write the path with the highest continuity cost into the safe path set. Then, perform consistency checks on the remaining paths with the first action of the safe path set. If the consistency check fails, delete the corresponding path. Output the safe path set corresponding to each scenario node.

[0010] In a preferred embodiment, the strategy solving module further includes: For each path in each set of safe paths, calculate the operating cost, which is the sum of the electricity purchase cost, hydrogen production electricity consumption cost, power generation consumption cost and state switching cost at each time. Then, perform a two-level sorting of the operating cost and the continuity cost of the corresponding path in the order of continuity cost first and operating cost second, and output the first-ranked path of each scenario node as the node path result. Read the node path results corresponding to each scenario node, perform cross-scenario statistics on hydrogen production, hydrogen storage, and power generation actions at the same time, extract the action combination with the most frequent occurrence as the common action at that time, and write the common action back to the first segment of the candidate path graph corresponding to each scenario node. Repeat the continuous cost calculation, operation cost calculation, and double-level sorting for each candidate path graph after writing it back until the common action is consistent in two consecutive rounds. Output the hydrogen production path, hydrogen storage path, and power generation path corresponding to each scenario node to form a scenario strategy set.

[0011] In a preferred embodiment, the cooperative execution module includes: The scenario paths in the scenario strategy set are expanded into action sequences in chronological order. The differences in hydrogen production, hydrogen storage, and power generation actions of each scenario path at each time are calculated. A scenario divergence matrix is ​​generated at each time. The scenario divergence matrix is ​​associated with the scenario probability value and oxygen supply gap value of the corresponding scenario node and written into the backtracking verification table. The scenario backtracking dataset is output. Based on the scenario backtracking dataset, a prefix consistency search is performed at each time point. The row sum of the scenario divergence matrix, the weighted divergence sum of the scenario probability values, and the weighted risk sum of the oxygen supply gap values ​​are calculated cumulatively at each time point. The continuous action segments with a cumulative result of zero are identified as consistent prefix segments, and the first action time with a non-zero cumulative result is identified as the conflict time. The consistent prefix segments and conflict times are output.

[0012] In a preferred embodiment, the cooperative execution module further includes: At the moment of conflict, calculate the joint cost value of each candidate action group corresponding to each scenario path, which consists of oxygen supply gap, hydrogen storage surplus deviation, electrolysis switching amount, and power generation switching amount. Sort the candidates according to the joint cost value and select the first candidate action group to write to the first conflict position of the current execution strategy. Write the first conflict position of the current execution strategy back to the corresponding time in the scenario backtracking dataset. Recalculate the hydrogen production action difference, hydrogen storage action difference, and power generation action difference for each scenario path after the moment of conflict. Repeat the prefix consistency search and joint cost value calculation until all action times in the current time period are written. Output the current execution strategy. Based on the hydrogen production, hydrogen storage, and power generation actions in the current execution strategy, electrolysis control commands, hydrogen storage control commands, and power generation control commands are generated respectively. These commands are then simultaneously sent to the electrolysis unit, hydrogen storage unit, and power generation unit. The corresponding electrolysis operation results, hydrogen storage status results, and power generation operation results are read, and the buffer execution results for the current time period are output.

[0013] In a preferred embodiment, the update module includes: Align the actual oxygen consumption results, actual photovoltaic results, actual electricity price results, and actual operation results corresponding to the buffer execution results with the scenario oxygen consumption results, scenario power supply results, scenario price results, and scenario operation results in the scenario strategy set according to time. Calculate the oxygen consumption deviation value, photovoltaic deviation value, electricity price deviation value, and operation deviation value respectively. Merge each deviation value according to the scenario path to form a scenario deviation sequence and output the deviation inversion result. Based on the deviation inversion results, the cumulative deviation value and path consistency value are calculated for each scenario path in the scenario policy set. The cumulative deviation value and path consistency value are written into the policy weights of each scenario path. The oxygen consumption deviation sequence, photovoltaic deviation sequence and electricity price deviation sequence in the deviation inversion results are written back to the temporal attention network. The updated policy weights are written back to the two-layer policy search, and the updated results are output. Based on the updated results, scenario prediction, policy search, and cooperative control are re-executed for the multi-source state sequence in the next time period, and the rolling buffer control results for the next time period are output.

[0014] The technical effects and advantages of this invention are as follows: 1. This solution constructs a scenario tree based on hospital oxygen consumption, photovoltaic power supply, electricity price and equipment operation data, and extracts the current execution strategy that is jointly established from multiple future scenarios, so that the current control no longer depends on a single prediction result, thereby relatively alleviating the problems of early hydrogen storage call, delayed electrolysis replenishment and inaccurate power generation switching caused by prediction deviation. 2. By first forming a multi-source state sequence, then extracting the joint features of oxygen consumption changes, photovoltaic available power supply changes, and electricity price changes, and generating candidate scenario nodes and correction scenario node sets accordingly, the evolutionary relationship in future time periods can be characterized in parallel, thus providing a more complete predictive basis for the subsequent collaborative solution of hydrogen production, hydrogen storage, and power generation. 3. By constructing a joint state point set and candidate path graph, and performing a two-level sorting based on continuity cost and operational cost, paths with oxygen supply gaps, hydrogen storage out-of-bounds, and large state jumps are first screened out, and then the node path results are determined, so that the obtained scenario strategy set can simultaneously take into account both oxygen supply continuity requirements and operational cost constraints. 4. Perform prefix consistency search and conflict resolution on each scenario path, extract cross-scenario common actions and recursively generate the current execution strategy, so that the electrolysis unit, hydrogen storage unit and power generation unit can execute common actions before the scenario is fully deployed, thereby relatively improving the adaptability of the current control strategy to the coexistence of multiple scenarios. 5. Align the actual oxygen consumption results, actual photovoltaic results, actual electricity price results, and actual operation results with the scenario results hourly, calculate the deviation inversion results and update the strategy weights, and then write them back to the prediction and solution process so that the control of the next period is based on the identified deviations, thereby forming a rolling correction mechanism and relatively improving the fit of subsequent scheduling. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system module structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figure 1 The present invention provides an electrolytic hydrogen production buffer system for medical oxygen supply, comprising a state construction module, a prediction module, a strategy solving module, a cooperative execution module, and an update module. The state construction module is used to acquire historical data on hospital oxygen consumption, weather forecasts, photovoltaic power generation forecasts, and grid electricity prices, as well as current electrolysis operation data, hydrogen storage status data, and power generation operation data. It performs time alignment, cross-source slicing, and feature encoding, and outputs a multi-source state sequence. In this embodiment, the state construction module organizes hospital oxygen consumption data, weather forecast data, photovoltaic power generation prediction data, grid electricity price data, and equipment operation data into a multi-source state sequence that the prediction module can directly access. Its processing logic does not involve directly splicing the raw data; instead, it first standardizes the time frame, then forms time slices based on changes in external driving conditions. Subsequently, it extracts state fields reflecting oxygen demand, photovoltaic power supply conditions, price changes, and equipment response relationships within each time slice. This ensures that the input objects read by the subsequent time-series attention network have a consistent time base, clear segment boundaries, and a fixed field structure. Historical time periods are used to form the multi-source state sequence, the current time period is used to write the current equipment state, and future and next time periods are not included in this process. Furthermore, in this embodiment, power supply-related fields correspond to the available photovoltaic power supply field, not the total power supply field. This implementation process includes the following steps: First, time mapping and gap filling under a unified time scale are performed on hospital oxygen consumption data, weather forecast data, photovoltaic power generation forecast data, grid electricity price data, electrolysis operation data, hydrogen storage status data, and power generation operation data to form an aligned dataset that can be directly compared and jointly calculated. The input includes the above seven types of raw data as well as the start and end times of the target time period. During processing, a unified time scale sequence is generated sequentially with a fixed step size, starting from the start time of the target time period. The fixed step size is the common step size of the hospital oxygen consumption data collection step size and the electrolysis operation data collection step size. Then, the raw timestamps of each data source are mapped to the corresponding time scale item by item. When multiple raw values ​​fall on the same time scale, the raw value with the shortest time distance between the timestamp and the time scale is retained. When there are no raw values ​​on the same time scale, if there are valid values ​​of the same field on both sides, the linear interpolation result is calculated according to the time distance ratio on both sides and written to the time scale. If there is only a valid value on one side, unilateral continuous filling is performed along the time direction. For meteorological forecast data and photovoltaic power generation prediction data, when the original source is given according to the forecast time and effective period, the corresponding field values ​​are written into the unified time scale covered according to the effective period. The output is an aligned dataset, which includes at least the time scale field, hospital oxygen consumption field, meteorological field, photovoltaic available power supply field, electricity price field, electrolysis operation field, hydrogen storage status field, power generation operation field, and missing measurement mark field. These are written into the oxygen consumption alignment table, power supply alignment table, price alignment table, and equipment operation alignment table, respectively, for unified reading in the subsequent slicing process. If a field has no original value in two or more consecutive unified time scales and cannot be filled by two-sided interpolation or one-sided continuation, then the field is recorded as a missing measurement field in the corresponding time scale and written into the missing measurement mark field. The time scale is still retained in the subsequent process, but the missing measurement field is not used as the basis for slicing. Subsequently, adjacent difference calculations were performed on the hospital oxygen value sequence, photovoltaic available power supply value sequence, and electricity value sequence in the aligned dataset. Time slices were formed based on the changes in external driving conditions to establish the correspondence between external driving change segments and equipment response segments. The input was the aligned dataset. During processing, the hospital oxygen value sequence, photovoltaic available power supply value sequence, and electricity value sequence were read in a unified time scale order, and the difference between two adjacent time scales was calculated hourly. When the previous difference and the next difference in the same sequence were positive and negative respectively, the next time scale was determined as the time when the difference sign changed. When any difference was zero, the nearest non-zero difference before the zero value was compared with the nearest non-zero difference after the zero value. When the two were positive and negative respectively, the time scale corresponding to the next non-zero difference was determined as the time when the difference sign changed. When the electricity values ​​of two adjacent time scales were different, the next time scale was determined as the time when the electricity value changed. When the last time scale of the historical period was connected to the first time scale of the current period, the connecting time scale was determined as the connection time. Then, all the times when the difference sign changes, the times when the electricity value changes, and the connection times are arranged in chronological order and deduplicated to form slice points. Time slices are then divided by the time interval between adjacent slice points. After the time slices are formed, the electrolysis operation data, hydrogen storage status data, and power generation operation data are written into the corresponding time slices according to their time assignments. This is because slice points are used to characterize the change boundaries of three types of external driving quantities: oxygen demand, photovoltaic power supply conditions, and electricity price conditions. Electrolysis, hydrogen storage, and power generation data belong to the equipment operation status that responds to external driving quantities and should be attached to the formed time slices. The output is a multi-source segment set, which includes at least the time slice number, start time scale, end time scale, intra-slice field set, and missing measurement marker field. These are written into the segment index table for the feature encoding process to read. If all external driving fields in a certain time slice are missing measurement fields, then that field will not participate in the intra-slice statistics and cross-source change correlation calculations for that time slice. However, the time slice is still retained, and the remaining valid fields and equipment operation fields continue to generate segment records. Finally, for each time slice in the multi-source segment set, intra-slice feature extraction, inter-slice difference calculation, and cross-source change correlation calculation are performed, and the state vector is formed by concatenating the data according to a fixed field order. The state vector is then arranged in chronological order to output a multi-source state sequence. The input is the multi-source segment set. During processing, for each time slice, the hospital oxygen consumption field, meteorological field, photovoltaic available power supply field, electricity price field, electrolysis operation field, hydrogen storage status field, and power generation operation field are read respectively. The slice start value, slice end value, intra-slice peak value, intra-slice valley value, and intra-slice mean value of each field are extracted as intra-slice features. Then, the slice start value of the corresponding field in the current time slice is subtracted from the slice end value of the same field in the previous time slice to form the inter-slice difference. Then, the product of the change in hospital oxygen consumption and the changes in photovoltaic available power supply, electricity price, electrolysis operation, hydrogen storage status, and power generation operation in the current time slice is calculated to form the cross-source change correlation value. The change in each field is the slice end value minus the slice start value. Then, the time slice identifier field, intra-slice feature field, inter-slice difference field, cross-source change association field, and missing test marker field are concatenated in a fixed field order to form the state vector of the corresponding time slice. The fixed field order is kept consistent in this embodiment to ensure that the reading position of the same type of field remains unchanged in the subsequent temporal attention network. The output is a multi-source state sequence, which is composed of state vectors arranged in chronological order and written into the input buffer of the prediction module for subsequent observation window division and temporal attention calculation. If a certain field has a missing test marker field in the current time slice, the intra-slice feature, inter-slice difference, and cross-source change association field corresponding to that field are all written with zero values, while the missing test marker field is retained to distinguish between the true zero value and the missing test written value. Through the above processing, the state construction module completes the standardized transformation from raw multi-source data to multi-source state sequences, enabling the input objects received by the prediction module to simultaneously possess a unified time caliber, clear slice boundaries, and fixed field structure. Furthermore, it establishes a correspondence between external driving quantities and device response quantities within the same time slice, thereby providing a consistent data foundation for solving subsequent oxygen consumption prediction results, photovoltaic available power supply prediction results, and price prediction results. In practical applications: At the start of each day's operation, the hospital oxygen supply platform first generates a unified time scale sequence covering historical and current time periods. It maps the hospital's oxygen consumption data from the previous operating cycle, the day's weather forecast data, photovoltaic power generation forecast data, electricity price time period data, and the current operating data of the electrolysis unit, hydrogen storage unit, and power generation unit to the unified time scale. Then, it identifies the locations of oxygen consumption changes caused by surgical scheduling, the locations of photovoltaic power supply changes caused by cloud cover, and the locations of time-of-use electricity price switching, and divides the time into time slices accordingly. The equipment operating status is written into each time slice according to its time assignment. Subsequently, the first value, last value, intra-slice statistical value, inter-slice difference, and cross-source change correlation value are extracted from each time slice. A multi-source state sequence is generated according to a fixed field order, and this multi-source state sequence is sent to the prediction module to support subsequent scenario tree construction, strategy solving, and buffer control.

[0018] The prediction module is used to input multi-source state sequences into a temporal attention network, perform joint extraction of oxygen consumption change features, photovoltaic fluctuation features, and electricity price change features, and construct a scenario tree for future periods based on the joint extraction results, outputting oxygen consumption prediction results, power supply prediction results, and price prediction results corresponding to each scenario node; In this embodiment, the prediction module's role is to form a scenario tree for future time periods based on the multi-source state sequence output by the state construction module, and further output the oxygen consumption prediction results, photovoltaic available power supply prediction results, and price prediction results corresponding to each scenario node. Its basic principle is as follows: First, within a continuous observation window, temporal features of hospital oxygen consumption changes, photovoltaic available power supply changes, and electricity price changes are extracted through temporal attention calculation, forming a shared state sequence that can be reused for subsequent scenario deduction. Then, the joint transition tensor for each future time period is constructed using the feature sequence and the shared state sequence, solving for the state distribution at each future time period and combining them to form candidate scenario nodes. Subsequently, unstable differentiation under the same parent node is eliminated through node cost calculation, sorting, field difference write-back, and repetition correction. Finally, edges are connected in chronological order, and the joint probability value is multiplied to generate the scenario tree. Here, the future time period is a time interval composed of several consecutive unified time scales after the current time period, and the next time period is the first unified time scale in the future time period. In this embodiment, the power supply state distribution and power supply prediction results correspond to the photovoltaic available power supply state distribution and photovoltaic available power supply prediction results. This implementation process includes the following steps: First, continuous observation window partitioning and temporal attention calculation are performed on the multi-source state sequence to form oxygen consumption characteristic sequences, photovoltaic characteristic sequences, electricity price characteristic sequences, and shared state sequences that can characterize the temporal dependencies of multiple sources. The inputs are the multi-source state sequence, the unified time scale sequence, and the observation window length. The observation window length is obtained by statistical estimation, taking the unified time scale number corresponding to three consecutive complete oxygen supply change cycles within the historical period, to ensure that the observation window simultaneously covers at least one oxygen consumption change segment, one photovoltaic change segment, and one electricity price change segment. During processing, the multi-source state sequence is first divided into consecutive observation windows in chronological order. Continue the observation window and add position encoding to the state vector in each observation window in sequence; then apply a causal mask to the state vector in the observation window so that the current time scale only receives information from the current time scale and the time scales before it; then read the hospital oxygen supply field, photovoltaic available power supply field, electricity price field and other operation fields in each state vector respectively, calculate the head attention score for the hospital oxygen supply field, photovoltaic available power supply field and electricity price field respectively. The head attention score is obtained by dividing the dot product of the query vector and the key vector by the square root of the dimension of the key vector, and then perform exponential normalization weighted summation to form three types of principal feature vectors; Then, cross-attention fusion is performed on the three main feature vectors and the auxiliary feature vectors corresponding to the remaining operating fields to form a shared state vector. The oxygen consumption feature sequence, photovoltaic feature sequence, electricity price feature sequence and shared state sequence are output in the order of observation windows. In the output, the shared state sequence includes at least the time scale field, oxygen consumption fusion field, photovoltaic fusion field, electricity price fusion field, as well as the electrolysis operation fusion field, hydrogen storage state fusion field and power generation operation fusion field, and is written into the joint transfer tensor to build a buffer for direct reading in the next step. If there is a missing measurement marker field in a certain observation window, the missing measurement marker field will also participate in the position encoding and causal mask calculation, so that the missing measurement information retains the source identifier in the attention calculation, instead of participating in the intra-head weighting with the real value. Subsequently, a joint transition tensor for each future time step is constructed based on the oxygen consumption feature sequence, photovoltaic feature sequence, electricity price feature sequence, and shared state sequence. The oxygen consumption state distribution, photovoltaic available power supply state distribution, and price state distribution are then solved to generate a candidate scenario node set and its joint probability value. The inputs are the oxygen consumption feature sequence, photovoltaic feature sequence, electricity price feature sequence, shared state sequence, and future time period length. The future time period length is taken as a number of consecutive unified time scales after the current time period. This number of unified time scales is converted from the time span corresponding to the system's rolling control cycle. During processing, a joint transition tensor is first constructed with future time step, oxygen consumption state, photovoltaic available power supply state, and price state as dimensions. Each tensor element is calculated by mapping the shared state vector of the corresponding future time step to the three types of feature vectors. Then, the joint transition tensor is accumulated along the non-target state dimension to obtain the oxygen consumption state distribution, photovoltaic available power supply state distribution, and price state distribution for the corresponding future time step. Subsequently, the state field values ​​from the three state distributions at the same future time are read and combined according to the time correspondence to form candidate scenario nodes. The distribution values ​​of the three state distributions on the candidate scenario node are multiplied to form a joint probability value. The output is the set of candidate scenario nodes and the joint probability value of each candidate scenario node. Each candidate scenario node includes at least a future time field, an oxygen consumption state field, a photovoltaic available power supply state field, a price state field, and a joint probability value field, and is written into the candidate node table for subsequent correction process. If there is a missing measurement propagation marker in any state distribution at a certain future time, the state field corresponding to the marker does not participate in the generation of candidate scenario nodes. The candidate scenario node is formed by combining the remaining valid state fields at the future time. After obtaining the candidate scene node set, the node cost value is calculated, sorted, field difference write-back, and duplicate correction is performed on each candidate scene node under the same parent node to eliminate unstable nodes in the scene differentiation process and form a corrected scene node set. The inputs are the candidate scene node set, joint probability value, shared state sequence, and parent node correspondence. During processing, the node cost value is calculated for each candidate scene node under the same parent node. The node cost value consists of three parts: the first part is the logarithmic sum of the joint probability values ​​of the candidate scene node; the second part is the sum of squared state transition differences between the corresponding state field of the candidate scene node and the corresponding state field of its parent node; and the third part is the sum of squared field reconstruction differences between the corresponding fields of the candidate scene node's state field after reverse mapping back to the shared state vector and the corresponding fields of the original shared state vector. Then, the nodes are sorted in ascending order of node cost value, and the first candidate scene node in the sort is retained as the provisional correction node under the current parent node. Then, for the remaining candidate scene nodes, calculate the difference vector between the corresponding fields of the first-ranked candidate scene node and the first-ranked candidate scene node. Write this difference vector back to the shared state vector corresponding to the future time when the candidate scene node was generated. The write-back positions are the oxygen fusion field, photovoltaic fusion field, and electricity price fusion field in the shared state vector. After the write-back is completed, re-execute the joint transition tensor construction, state distribution solution, candidate scene node generation, and node cost calculation. The output is the corrected scene node set. The stopping condition is that the first-ranked candidate scene node under the same parent node is consistent in two consecutive rounds, and the corresponding field difference vectors are consistent item by item. This stopping condition means that the scene differentiation result and the write-back correction amount under the current parent node no longer change. Therefore, the correction ends and the corrected scene node set is output. If there is only one candidate scene node under a certain parent node, the candidate scene node is directly written into the corrected scene node set without executing the write-back loop. Finally, edge connections and probability accumulation are performed on the set of corrected scene nodes in the order of future time moments to generate a scene tree. The state field values ​​in each scene node are read to form oxygen consumption prediction results, photovoltaic available power supply prediction results, and price prediction results. The input quantities are the set of corrected scene nodes, the correspondence between parent nodes, and the joint probability value. During processing, each corrected scene node is first connected to its corresponding parent node in the previous future time moment in the order of future time moments to form a directed hierarchical structure from the end of the current time period to the future time period. Then, the joint probability value of each connection is multiplied layer by layer to obtain the path probability value from the root node to the current scene node. Then, the path probability values ​​of all scene nodes under the same future time moment are normalized to obtain the standard probability value of the same layer, and the standard probability value of the same layer is written to the corresponding scene node. After the scene tree is generated, the oxygen consumption state distribution field value is read from each scene node to form the oxygen consumption prediction result, the photovoltaic available power supply state distribution field value is read to form the photovoltaic available power supply prediction result, and the price state distribution field value is read to form the price prediction result. In the output, the oxygen consumption prediction result includes at least the prediction time field, the scene node identifier field, and the predicted oxygen consumption field; the photovoltaic available power supply prediction result includes at least the prediction time field, the scene node identifier field, and the predicted power supply field; and the price prediction result includes at least the prediction time field, the scene node identifier field, and the predicted electricity value field. The above results are written together into the scene tree result table for the strategy solving module to read by scene node. If there is no connectable parent node at the same future time, an edge is established between the corrected scene node at that time and the scene node with the highest path probability value in the previous future time to ensure the continuity of the scene tree in the time dimension. Through the above processing, the prediction module completes the transformation from multi-source state sequences to scene trees and prediction results of each scene node. This enables the subsequent strategy solving module to simultaneously read future oxygen consumption results, future photovoltaic power supply results, and future price results under the same scene node, and organize the solutions for hydrogen production paths, hydrogen storage paths, and power generation paths with a unified scene identifier. This process also supplements the field composition of the shared state sequence, the structure of the joint transition tensor, the solution method of the state distribution, the write-back position and function of field differences, and the meaning of the stopping conditions of the prediction module, thereby eliminating the problems of unclear object boundaries and suspended computational connections. In practical applications: After completing state construction, the hospital oxygen supply platform feeds the multi-source state sequence into a temporal attention network according to a continuous observation window, obtaining three types of feature sequences reflecting changes in oxygen consumption during surgical shifts, changes in available photovoltaic power supply caused by cloud cover, and changes in time-of-use electricity prices, as well as a shared state sequence. Subsequently, a joint transition tensor is constructed based on several unified future time scales to form candidate scenario nodes for multiple future moments. Node cost calculations and field difference write-backs are performed on candidate scenario nodes under the same parent node until the scenario differentiation results stabilize. Finally, the stabilized corrected scenario nodes are sequentially connected to generate a scenario tree, and the oxygen consumption prediction results, available photovoltaic power supply prediction results, and price prediction results corresponding to each scenario node are output for direct use in subsequent two-layer strategy searches.

[0019] The strategy solving module, based on the oxygen consumption prediction results, power supply prediction results, and price prediction results corresponding to each scenario node, combined with electrolysis operation data, hydrogen storage status data, and power generation operation data, performs a two-level strategy search with the objectives of prioritizing oxygen supply continuity and operating cost constraints, solves the hydrogen production path, hydrogen storage path, and power generation path corresponding to each scenario node, and outputs the scenario strategy set. In this embodiment, the role of the strategy solving module is to solve the hydrogen production path, hydrogen storage path and power generation path corresponding to each scenario node based on the oxygen consumption prediction results, photovoltaic available power supply prediction results and price prediction results corresponding to each scenario node output by the prediction module, and combined with the electrolysis operation data, hydrogen storage status data and power generation operation data of the current period. The basic principle is as follows: First, the predicted results of the scenario nodes are aligned with the operating status of the equipment at the same time reference, and a set of connectable joint state points is discretized. Then, a candidate path graph is constructed according to the equipment adjustment constraints, hydrogen storage constraints, and oxygen and power supply balance constraints. Subsequently, prioritizing oxygen supply continuity, a set of safe paths is screened out using continuity costs. Then, a second-level sorting is completed within the set of safe paths using operating costs to obtain the node path results corresponding to each scenario node. Finally, cross-scenario statistics and common action write-back are performed on the node path results of each scenario node to fix the actions that can be jointly executed in the current time period, thereby forming a scenario strategy set. Here, the power supply prediction results correspond to the photovoltaic available power supply prediction results. The electrolysis power, hydrogen storage flow, and power generation in the joint state point set are all discretized according to the allowable adjustment steps of the equipment. The current time period is the time period in which the strategy is implemented, and the future time period is used for path expansion. The implementation process includes the following steps: First, the oxygen consumption forecast, photovoltaic power supply forecast, and price forecast results corresponding to each scenario node are aligned with the electrolysis operation data, hydrogen storage status data, and power generation operation data by time. Then, a candidate path graph is constructed for each scenario node. The purpose is to form a searchable path space that satisfies the equipment adjustment range and the oxygen and power supply balance. Inputs include the forecast results for each scenario node in the scenario tree result table, the current time period's electrolysis operation data, hydrogen storage status data, power generation operation data, and the equipment's allowable adjustment steps, including the electrolysis power adjustment step, hydrogen storage flow rate adjustment step, and power generation step. The rate adjustment step is given by the constraints of the equipment operation rules, and the adjacent adjustment amounts allowed by the corresponding equipment control interface are taken respectively. During processing, the oxygen consumption prediction results, photovoltaic available power supply prediction results, and price prediction results corresponding to each scenario node are first aligned with the current electrolysis operation data, hydrogen storage status data, and power generation operation data hourly according to a unified time scale. Then, taking the current measured electrolysis power, hydrogen storage flow rate, and power generation power as the starting point, the discrete state is expanded to the future time according to their respective adjustment steps to form a joint state point set composed of the discrete values ​​of electrolysis power, hydrogen storage flow rate, and power generation power. Then, edges are established between adjacent joint state points. The edge conditions include: the differences between the electrolysis power, hydrogen storage flow rate, and power generation at the joint state point at the next time step and their corresponding values ​​at the previous time step fall within their respective allowable adjustment step ranges; the hydrogen storage capacity at the next time step is obtained by multiplying the hydrogen storage capacity at the previous time step by the hydrogen storage flow rate at the previous time step and the time step; and the hydrogen storage capacity at the next time step is between the upper and lower bounds recorded by the hydrogen storage unit. Simultaneously, the sum of the available photovoltaic power supply, purchased electricity, and power generation at the next time step, together with the electrolysis power consumption, satisfies the oxygen supply and power balance relationship at that time step. The oxygen supply and power balance relationship is determined by the predicted oxygen consumption of the hospital. The balance between the equivalent amount of hydrogen production and oxygen supply from electrolysis and the equivalent amount of hydrogen storage and oxygen release; the output is the candidate path graph corresponding to each scenario node. The candidate path graph includes at least the path start point, path end point, joint state points at each time step, edge relationships, and corresponding scenario node identifiers, and is written into the path graph buffer for the next continuous cost calculation; if a scenario node cannot form a subsequent joint state point that satisfies the edge conditions at a certain future time step, the relevant edge passing through that time step is deleted; if a scenario node has no connected path at all future time steps, the single-step path formed by the current action is written into the candidate path graph for subsequent safety verification; Subsequently, the continuity cost of each candidate path is calculated, and the set of safe paths corresponding to each scenario node is screened out. The purpose is to first eliminate paths that may cause oxygen supply interruption, hydrogen storage overrun, or excessive equipment jumps. The input quantities are the candidate path graph corresponding to each scenario node, the upper and lower bounds of hydrogen storage, the predicted oxygen consumption of the hospital at each time, and the equipment state sequence corresponding to each path. During processing, the joint state points in each path are read hourly, and the oxygen supply gap of the path at each time is calculated. The oxygen supply gap is the sum of the predicted oxygen consumption of the hospital at that time minus the equivalent amount of electrolytic oxygen supply and the equivalent amount of hydrogen storage release oxygen supply corresponding to the path. Then, the amount of hydrogen storage overrun at each time is calculated. The amount of hydrogen storage overrun is the difference between the hydrogen storage at that time and the lower bound of hydrogen storage or the upper bound of hydrogen storage. If there is no overrun, it is recorded as zero. Then, the state jump variable between the joint state points at each adjacent time is calculated. The state jump variable is the sum of the square of the difference in electrolytic power, the square of the difference in hydrogen storage flow, and the square of the difference in power generation at adjacent time. Then, the sum of squares of the oxygen supply gap, the sum of squares of the hydrogen storage overrun, and the sum of squares of the state jump variables at each time step are added to form the continuity cost of the corresponding path. All paths under the same scenario node are then sorted and retained one by one according to their continuity costs from smallest to largest. The first path in the sorted list is directly written into the safe path set. The remaining paths are then checked for consistency with the retained paths in the safe path set, where the first action is the combination of hydrogen production, hydrogen storage, and power generation actions corresponding to the first future time step of the path. If the consistency check is successful, the path is retained; otherwise, it is deleted. The output is the safe path set corresponding to each scenario node, and the continuity cost of each safe path is written into the path record table for the next step of cost calculation. If multiple paths have the same continuity cost, the path with the smaller oxygen supply gap at the first future time step is ranked first. If the values ​​are still the same, the path with the smaller hydrogen storage overrun at the first future time step is ranked first. After obtaining the set of safe paths, the operating cost of each path in the set is calculated, and a two-level sorting is performed in the order of continuity cost first and operating cost second to determine the node path results corresponding to each scenario node. The purpose is to further reduce the operating cost while meeting the oxygen supply continuity requirement. The inputs are the set of safe paths corresponding to each scenario node, the price prediction results at each time, the electrolysis power consumption conversion relationship, the power generation consumption conversion relationship, and the state switching record. During processing, the electricity purchase cost, hydrogen production power consumption cost, power generation consumption cost, and state switching cost are calculated hourly for each path. The electricity purchase cost is the product of the electricity purchase amount at that time and the price prediction value at that time. The hydrogen production power consumption cost is the product of the electrolysis power at that time, the time step, and the corresponding price prediction value. The power generation consumption cost is the product of the power generation power at that time, the time step, and the power generation unit consumption conversion coefficient. The state switching cost is the sum of the switching costs corresponding to the difference in electrolysis power, the difference in hydrogen storage flow, and the difference in power generation power between adjacent times. The costs at each time point are then summed to form the operating cost of the corresponding path. Subsequently, all safe paths under the same scenario node are first sorted by continuity cost, and then among paths with the same continuity cost, they are sorted by operating cost. The first path in this two-level sorting is determined as the node path result corresponding to that scenario node, and the node path result is written into the node path table. The output is the node path result corresponding to each scenario node, where each node path result includes at least the scenario node identifier, path sequence, continuity cost, and operating cost. If the continuity cost and operating cost of multiple paths are the same, the path with the smaller state switching cost at the first future time point is selected as the node path result. If this value is still the same, the path with the higher final hydrogen storage value is selected as the node path result to retain subsequent buffer margins. Finally, the node path results corresponding to each scenario node are read, and cross-scenario statistics are performed on hydrogen production, hydrogen storage, and power generation actions at the same time. The hydrogen production path, hydrogen storage path, and power generation path corresponding to each scenario node are formed by writing back the common actions. The purpose is to extract the actions that can be jointly executed in the current time period from the node path results of different scenario nodes, so as to avoid the action dispersion in the execution layer before the scenario is fully unfolded. The input is the node path table, the candidate path diagram, and the node identifier of each scenario. During processing, the hydrogen production, hydrogen storage, and power generation actions corresponding to the time of the corresponding time in all node path results are read hourly according to the future time. Cross-scenario statistics are performed on the action combinations at the same time, and the action combination with the most occurrences is taken as the common action at that time. When multiple action combinations have the same number of occurrences, the action combination with the smaller cumulative value of continuous cost is selected. If they are still the same, the action combination with the smaller cumulative value of running cost is selected. Then, the common actions at that moment are written back to the first segment of the candidate path graph corresponding to each scenario node, and the continuity cost calculation, running cost calculation, and two-level sorting are re-executed accordingly. If the common actions extracted hourly in two consecutive rounds are completely consistent, the write-back loop is stopped, and the node path results of each scenario node in the stopped round are read. The corresponding hydrogen production path, hydrogen storage path, and power generation path are output respectively to form a scenario strategy set. The output is the scenario strategy set, which includes at least the scenario node identifier, hydrogen production path, hydrogen storage path, power generation path, and common action record, and is written to the collaborative execution module call table. If there is only one action combination in the path results of all nodes at a certain moment, the action combination is directly recorded as a common action without parallel resolution. If there is no available path in the candidate path graph corresponding to a scenario node after the common action is written back, the node path results of the previous round of write-back for that scenario node remain unchanged, and it continues to participate in the common action statistics of the next moment. Through the above processing, the strategy solving module completes the construction process from the prediction results of the scenario nodes to the scenario strategy set. This ensures that each scenario node not only has future oxygen consumption results, photovoltaic power supply results, and price results, but also obtains hydrogen production paths, hydrogen storage paths, and power generation paths that meet the constraints of equipment regulation, hydrogen storage capacity, and oxygen and power supply balance. This process also supplements the discretization rules of the joint state point set, the adjacent edge constraints, the calculation caliber of hydrogen storage out-of-bounds amount, the consistency verification object of the first action, the parallel resolution rules of common actions, and the scope of action of common actions, thereby eliminating the problems of unclear search space, sorting basis, and stopping rules. In practical applications: After reading the scenario tree results, the hospital oxygen supply platform aligns the oxygen consumption prediction results, photovoltaic power supply prediction results, and price prediction results corresponding to each scenario node within a certain future time period with the current operating status of the electrolysis unit, hydrogen storage unit, and power generation unit. It then unfolds a joint state point set of electrolysis power, hydrogen storage flow rate, and power generation power according to the allowable adjustment steps of the equipment, and generates candidate path maps corresponding to each scenario node. Subsequently, it first calculates the continuity cost based on the oxygen supply gap, hydrogen storage overrun, and state jump variables, and filters out safe paths. Then, it calculates the operating cost based on the electricity purchase cost, hydrogen production power consumption cost, power generation consumption cost, and state switching cost, and determines the node path results for each scenario node. Finally, it performs cross-scenario statistics on the action combinations at the same time, extracts common actions, and writes them back until two consecutive rounds of common actions are consistent, obtaining a scenario strategy set that can be directly called by the collaborative execution module.

[0020] The collaborative execution module is used to perform scenario backtracking consistency verification and cross-scenario common prefix extraction on the scenario strategy set, generate the current execution strategy, and perform collaborative control on the electrolysis unit, hydrogen storage unit and power generation unit according to the current execution strategy, and output the buffer execution results of the current time period; In this embodiment, the collaborative execution module extracts the directly executable actions from the scenario strategy set for the current time period, transforms these actions into collaborative control commands for the electrolysis unit, hydrogen storage unit, and power generation unit, and outputs the buffered execution results for the current time period. Its basic principle is as follows: First, each scenario path is expanded into a sequence of actions arranged chronologically, and a scenario backtracking dataset is constructed using the action differences between scenarios, scenario probability values, and oxygen supply gap values. Then, a prefix consistency search is performed on the scenario backtracking dataset to determine the consistent prefix segments that can be jointly executed and the conflict moments that need to be resolved. Subsequently, the joint cost value is calculated for candidate action groups within the conflict moments and recursively written into the current execution strategy until all action moments for the current time period are written. Finally, control commands are generated based on the current execution strategy and sent synchronously, so that the buffered control results directly correspond to the continuity of medical oxygen supply. Here, the scenario probability value comes from the standard probability value of the corresponding scenario node in the scenario tree, the oxygen supply gap value comes from the oxygen supply gap amount at the corresponding moment for each scenario path in the scenario strategy set, and the rows and columns of the scenario divergence matrix correspond to the scenario paths. This implementation process includes the following steps: First, each scenario path in the scenario strategy set is expanded into an action sequence in chronological order, and a scenario backtracking dataset is constructed to form a unified comparison basis required for subsequent consistency search and conflict resolution. The input includes the scenario strategy set, the scenario probability values ​​in the scenario tree result table, and the oxygen supply gap value at the corresponding time for each scenario path. During processing, each scenario path is first expanded into an action sequence according to the unified time scale of the current period, where each action sequence element includes the hydrogen production action, hydrogen storage action, and power generation action at that time. Then, for any two scenario paths at the same time, the difference in hydrogen production action, hydrogen storage action, and power generation action are calculated respectively, and the three are concatenated in a fixed order to form corresponding matrix elements, thereby generating a scenario divergence matrix at each time, where the rows and columns of the matrix correspond to the scenario paths, and the matrix elements correspond to the action difference combination values ​​of the two scenario paths at that time. Subsequently, the scenario divergence matrix at each moment is associated with the scenario probability value and oxygen supply gap value of the corresponding scenario path and written into the backtracking verification table. The backtracking verification table includes at least the time scale field, scenario path identifier field, scenario divergence matrix field, scenario probability value field, and oxygen supply gap value field. The output is the scenario backtracking dataset, which is written into the collaborative execution cache for the next step of prefix consistency search. If there is only one scenario path at a certain moment, the scenario divergence matrix at that moment is recorded as a cell matrix, and the matrix elements are written as zero values. At the same time, the scenario probability value and oxygen supply gap value records at that moment are retained. Subsequently, based on the scene backtracking dataset, a prefix consistency search is performed at each time point to determine the consistent prefix segments that can be executed jointly within the current time period and the position of the first action that needs conflict resolution. The input is the scene backtracking dataset. During processing, the scene divergence matrix, scene probability value, and oxygen supply gap value are read hourly in chronological order. Then, the scene divergence matrix at the current time is summed row by row to obtain the cumulative action divergence amount corresponding to each scene path. Then, the cumulative action divergence amount is weighted by the scene probability value corresponding to each scene path to obtain the weighted divergence sum at that time. Finally, the cumulative action divergence amount is weighted by the oxygen supply gap value corresponding to each scene path to obtain the weighted risk sum at that time. To ensure a unified calculation standard for the aforementioned cumulative values, before writing them into the backtracking verification table, the differences in hydrogen production, hydrogen storage, and power generation actions are first converted into a unified action difference scalar value according to a fixed field order. Then, this unified action difference scalar value is used to form matrix elements. Therefore, subsequent row sums, weighted divergence sums, and weighted risk sums can be accumulated hourly under the same scalar standard. Then, starting from the beginning of the current time period, the row sums, weighted divergence sums, and weighted risk sums for each time period are accumulated in chronological order. When the cumulative result remains zero, the corresponding continuous action segment is determined as a consistent prefix segment. When the cumulative result first shows a non-zero value, the corresponding time scale is determined as a conflict moment. The output is the consistent prefix segment and the conflict moment, which are written into the current execution strategy initialization table. If the cumulative result for all times within the current time period is zero, the entire current time period is directly determined as a consistent prefix segment, and the conflict moment is recorded as empty. The current execution strategy is then directly generated from the consistent prefix segment. After obtaining the conflict time, the joint cost value calculation, sorting, and recursive writing of the candidate action groups corresponding to each scenario path at the conflict time are performed to form the current execution strategy hourly. The input includes the consistent prefix segment, the conflict time, the scenario backtracking dataset, and the candidate action groups corresponding to each scenario path at the conflict time. During processing, the hydrogen production action, hydrogen storage action, and power generation action corresponding to each scenario path at the conflict time are read first to form candidate action groups. Then, the joint cost value is calculated for each candidate action group. The joint cost value consists of four parts: the oxygen supply gap corresponding to the conflict time, the deviation between the hydrogen storage surplus formed after executing the candidate action group at the conflict time and the average hydrogen storage surplus of the current scenario path, the electrolysis switching amount and the power generation switching amount of the candidate action group relative to the action executed at the previous time. The above four items are added in a fixed field order to form the joint cost value of the corresponding candidate action group. Then, the action groups are sorted in ascending order of their joint value, and the first candidate action group is selected and written into the first conflict position of the current execution strategy. After writing, the hydrogen production, hydrogen storage, and power generation actions corresponding to the first conflict position are written back to the corresponding time in the scenario backtracking dataset, and the first action reference value of each scenario path after the conflict time is replaced with the write-back result. Then, the differences in hydrogen production, hydrogen storage, and power generation actions are recalculated for each scenario path after the conflict time, and the scenario divergence matrix for subsequent time periods is regenerated. Prefix consistency search and joint value calculation are performed again. In each round of this recursive process, only the action group corresponding to the current first conflict time is written. After writing, the scenario divergence relationship for subsequent time periods is immediately recalculated until all action time periods in the current time period have been written. The output is the current execution strategy, which includes at least a time scale field, a hydrogen production action field, a hydrogen storage action field, and a power generation action field, and is written into the execution control table for the generation of the next instruction. If the joint value of multiple candidate action groups is the same at a certain conflict time, the candidate action group with the higher probability value of the corresponding scenario is selected. If the value is still the same, the candidate action group with the smaller oxygen supply gap is selected. Finally, control commands are generated based on the hydrogen production, hydrogen storage, and power generation actions in the current execution strategy, and simultaneously sent to the electrolysis unit, hydrogen storage unit, and power generation unit to obtain the buffered execution results for the current time period. The input quantities are the current execution strategy and the control interface parameters of each device. During processing, the hydrogen production action field, hydrogen storage action field, and power generation action field in the current execution strategy are read hourly and converted into electrolysis control commands, hydrogen storage control commands, and power generation control commands, respectively. The electrolysis control command includes at least a target electrolysis power field and an execution time field, the hydrogen storage control command includes at least a target hydrogen storage flow field and an execution time field, and the power generation control command includes at least a target power generation field and an execution time field. Then, the three types of control commands are simultaneously sent to the electrolysis unit, hydrogen storage unit, and power generation unit at the same execution time. After execution, the equipment reads the electrolysis operation results, hydrogen storage status results, and power generation operation results respectively. The electrolysis operation results include at least the measured electrolysis power and the corresponding time; the hydrogen storage status results include at least the measured hydrogen storage volume, the measured hydrogen storage flow rate, and the corresponding time; and the power generation operation results include at least the measured power generation power and the corresponding time. Then, the three types of results are associated with the corresponding time of the current execution strategy and written into the buffer execution result table. The output is the buffer execution result for the current time period, which includes at least the execution time field, the hydrogen production execution result field, the hydrogen storage execution result field, the power generation execution result field, and the corresponding oxygen supply gap compensation field. This result is written into the update module call table for subsequent deviation inversion and strategy weight updates. If any equipment does not return a measured result at a certain execution time, the control command corresponding to that execution time is recorded as a non-response command. Simultaneously, the most recently returned measured result is written into the corresponding result field, and a non-response mark is written into the buffer execution result table. Through the above processing, the collaborative execution module completes the transformation from the scenario strategy set to the current execution strategy and then to the buffered execution result. This enables the system to prioritize the extraction of actions that can be jointly executed in the current period when multiple future scenarios exist simultaneously, and to gradually resolve path divergences at the moment of conflict, ultimately forming a collaborative control result that directly corresponds to the continuity of medical oxygen supply. This process also supplements the dimensions of the scenario divergence matrix, the source of scenario probability values ​​and oxygen supply gap values, the unified calculation caliber of cumulative results, and the recursive writing logic of the first conflict position, thereby eliminating the problems of unclear object boundaries and unclear recursive relationships in backtracking verification and execution control. In practical applications: After obtaining the scenario strategy set, the hospital oxygen supply platform first expands each scenario path into an action sequence covering the current time period, and generates a scenario divergence matrix hourly. The standard probability value and oxygen supply gap value of the corresponding scenario at the same level are written into the backtracking verification table. Then, the consistency search identifies the action segments that can be jointly executed in the first half of the current time period and locates the first conflict moment. Subsequently, at the conflict moment, the joint cost value of the candidate action group is calculated and the first-ranked action group is selected and written into the current execution strategy. After writing, it continues to push forward until all action moments of the current time period are determined. Finally, the formed hydrogen production action, hydrogen storage action, and power generation action are synchronously sent to the electrolysis unit, hydrogen storage unit, and power generation unit. The equipment results after execution are read and buffered execution results are formed for the deviation inversion of the subsequent update module.

[0021] The update module is used to obtain the actual oxygen consumption results, actual photovoltaic results, actual electricity price results and actual operation results corresponding to the buffer execution results, perform deviation inversion and policy weight update with the scenario policy set, and write the update results back to the temporal attention network and the two-layer policy search, and output the rolling buffer control results for the next time period. In this implementation, the update module compares the buffer execution results of the current time period with the scene results corresponding to the scene policy set item by item, inverts the sources of deviation in the current prediction and solution chain, and writes the deviation results and policy weights back to the prediction module and the policy solution module to output the rolling buffer control results for the next time period. Its basic principle is: first, align the actual results with the scene results under a unified time scale to form deviation inversion results that correspond to scene paths; then, calculate the cumulative deviation value and path consistency value of each scene path based on the deviation inversion results, and update the policy weights accordingly. Simultaneously, write the oxygen deviation sequence, photovoltaic deviation sequence, and electricity price deviation sequence into the input field of the time-series attention network, and write the updated policy weights into the two-layer policy search process; finally, re-execute scene prediction, policy search, and collaborative control for the next time period based on the update results. Here, the scene power supply result corresponds to the photovoltaic available power supply result, and the scene operation result consists of the hydrogen production execution volume, hydrogen storage execution volume, and power generation execution volume corresponding to each scene path at each time in the scene policy set. The next time period is the first unified time scale in the future time period. This implementation process includes the following steps: First, the actual oxygen consumption results, actual photovoltaic results, actual electricity price results, and actual operation results corresponding to the buffer execution results are aligned with the scenario oxygen consumption results, scenario power supply results, scenario price results, and scenario operation results in the scenario strategy set according to time. Then, oxygen consumption deviation values, photovoltaic deviation values, electricity price deviation values, and operation deviation values ​​are calculated separately to form deviation inversion results. Inputs include the buffer execution result table, actual oxygen consumption records, actual photovoltaic records, actual electricity price records, the scenario strategy set, and a unified time scale sequence. During processing, the actual oxygen consumption results are first aligned with the scenario oxygen consumption results hourly according to the unified time scale. The actual photovoltaic results are aligned with the scenario power supply results hourly, the actual electricity price results are aligned with the scenario price results hourly, and the actual operation results are aligned with the scenario operation results hourly. Then, the oxygen consumption deviation value, photovoltaic deviation value, electricity price deviation value, and operation deviation value are calculated separately. The oxygen consumption deviation value is the actual oxygen consumption result minus the scenario oxygen consumption result, the photovoltaic deviation value is the actual photovoltaic result minus the scenario power supply result, the electricity price deviation value is the actual electricity price result minus the scenario price result, and the operation deviation value is the combination of the differences between the hydrogen production execution volume, hydrogen storage execution volume, and power generation execution volume in the actual operation results and the corresponding quantities in the scenario operation results. Subsequently, the deviation values ​​are merged according to the scenario path to form a scenario deviation sequence for the corresponding scenario path. Each scenario deviation sequence includes at least a time scale field, an oxygen consumption deviation field, a photovoltaic deviation field, an electricity price deviation field, and an operation deviation field. The output is the deviation inversion result, which is written to the deviation result table for subsequent strategy weight updates. If the actual result is missing at a certain moment, the most recent returned actual result at that moment is read and one-sided continuation filling is performed, and a missing test mark is written to the deviation result table. If a scenario path does not have a corresponding scenario result at that moment, the deviation field of that scenario path at that moment is recorded as zero and a no-scenario mark is written. Subsequently, based on the deviation inversion results, the cumulative deviation value and path consistency value are calculated for each scenario path in the scenario policy set. The cumulative deviation value and path consistency value are then written into the policy weights of each scenario path. At the same time, the time-series attention network and the two-layer policy search are written back to form the updated results. The inputs are the deviation inversion results, the scenario policy set, and the current policy weights of each scenario path. The current policy weights are initialized with the same initial value when each scenario path is first generated and written into the path record table. During processing, the scenario deviation sequence of each scenario path is first accumulated in chronological order by adding the oxygen consumption deviation value, photovoltaic deviation value, electricity price deviation value, and operation deviation value to obtain the cumulative deviation value of the scenario path. Then, the action sequence of the scenario path is compared hourly with the actual action sequence executed in the current time period. The number of times when hydrogen production, hydrogen storage, and power generation actions are simultaneously consistent is counted. The number of consistent times is divided by the total number of times in the current time period to obtain the path consistency value of the corresponding scenario path. Subsequently, the cumulative deviation value and path consistency value are written into the policy weight record of each scenario path in a fixed field order. The cumulative deviation value is used to characterize the degree of deviation between the scenario path and the actual execution result, and the path consistency value is used to characterize the degree of consistency between the scenario path and the actual execution action. In this embodiment, the policy weight is updated to the absolute value of the path consistency value minus the cumulative deviation value. Then, the oxygen deviation sequence, photovoltaic deviation sequence, and electricity price deviation sequence in the deviation inversion result are written into the additional fields of the time slice corresponding to the next round of multi-source state sequence as one of the input fields of the temporal attention network. The updated policy weight is written into the record corresponding to each scenario path as an additional quantity for path ranking in the next round of two-layer policy search. The output is the update result, which includes at least the cumulative deviation value, path consistency value, updated policy weight, and the deviation additional field after writing back, and is written into the prediction module update table and the policy solution update table respectively. If the updated policy weights of multiple scenario paths are the same, the scenario path with the larger path consistency value is retained first; if the value is still the same, the scenario path with the smaller cumulative deviation value is retained first. Finally, based on the updated results, scene prediction, policy search, and collaborative control are re-executed on the multi-source state sequence for the next time period to output the rolling buffer control result for the next time period. The inputs are the updated results, the starting time scale of the next time period, and the multi-source state sequence corresponding to the next time period. During processing, the deviation additional field written to the update table of the prediction module is read first and concatenated with the multi-source state sequence of the next time period to form the updated input state sequence. Then, the updated input state sequence is input into the temporal attention network to re-execute the scene tree construction and output the scene node prediction result corresponding to the next time period. Subsequently, the updated policy weights written to the policy solution update table are read and used as additional quantities for path sorting in the two-layer policy search. The joint state point set construction, continuous cost calculation, running cost calculation, and scene policy set generation are re-executed. Then, based on the updated scenario strategy set, the prefix consistency search, conflict resolution, and control command issuance are re-executed, and the rolling buffer control results for the next time period are output. The rolling buffer control results include at least the execution time field for the next time period, the hydrogen production control result field, the hydrogen storage control result field, the power generation control result field, the oxygen supply gap compensation field, and the corresponding scenario path identifier field, and are written into the next round of buffer execution result table for subsequent deviation inversion. If the actual result of the last unified time scale of the current time period has not been received before the start of the next time period, the results that have been updated before the unified time scale are retained for rolling control, and the deviation field corresponding to the actual result is added in the next round of update. Through the above processing, the update module completes the full update from the current time period buffer execution result to the next time period rolling buffer control result, enabling the system to promptly identify prediction deviation, running deviation, and path deviation after each round of execution, and directly apply the deviation to the next round of scenario prediction and path sorting; this process also supplements the field composition of the scenario running result, the calculation caliber of the cumulative deviation value and the path consistency value, the write-back object of the deviation sequence and strategy weight, and the next round call relationship, thereby eliminating the problems of unclear object destination, unclear weight purpose, and insufficient rolling control connection in the update module; In practical applications: After the current time period ends, the hospital oxygen supply platform first compares the actual oxygen consumption, photovoltaic power generation, electricity price, and equipment operation results of that time period with the scenario results corresponding to the scenario strategy set on a uniform time scale to form the deviation inversion results of each scenario path; then it calculates the cumulative deviation value and path consistency value of each scenario path, updates the strategy weights, and writes the oxygen consumption deviation sequence, photovoltaic power generation deviation sequence, and electricity price deviation sequence into the additional fields of the next round of multi-source state sequence; then it sends the updated input state sequence into the prediction module and the updated strategy weights into the strategy solving module to regenerate the scenario tree, scenario strategy set, and current execution strategy for the next time period, and finally outputs the rolling buffer control results for the next time period.

[0022] Working Principle: This solution first uses a state construction module to unify hospital oxygen consumption data, weather forecast data, photovoltaic power generation prediction data, grid electricity price data, and operational data from electrolysis, hydrogen storage, and power generation under the same time base. This data is then segmented into time slices that reflect stages of external change and encoded into a multi-source state sequence. The prediction module then extracts features of oxygen consumption changes, photovoltaic available power supply changes, and electricity price changes from the multi-source state sequence to construct a scenario tree for future time periods, obtaining the oxygen consumption prediction results, photovoltaic available power supply prediction results, and price prediction results for each scenario node. Finally, the strategy solving module, based on the prediction results of each scenario node and the current equipment status, solves for the hydrogen production corresponding to each scenario node. The system identifies pathways, hydrogen storage pathways, and power generation pathways, forming a scenario strategy set. The collaborative execution module then extracts actions that can be jointly executed in the current time period from multiple scenario pathways, generates the current execution strategy, and distributes it to the electrolysis unit, hydrogen storage unit, and power generation unit to obtain buffered execution results. Finally, the update module compares the actual results with the scenario results, inverts the deviation, updates the strategy weights, and writes them back to the prediction and solution process to form the rolling buffer control results for the next time period. Therefore, the entire system forms a complete control chain of state construction, scenario prediction, strategy solution, collaborative execution, and rolling update. The core is to always organize the collaborative buffer between hydrogen production, hydrogen storage, and power generation around the continuity of medical oxygen supply. For example, during the daytime operation of a hospital, the increased surgical schedule in the morning leads to a gradual increase in oxygen demand, while cloud cover at noon reduces the available power supply from photovoltaic systems. Simultaneously, peak electricity prices arrive. The system first organizes these multi-source changes into a unified state sequence, then predicts the changes in oxygen consumption, photovoltaic power, and electricity prices under different scenarios in the future. Subsequently, for each scenario, it calculates the corresponding hydrogen production, hydrogen storage, and power generation paths, and extracts the most suitable common actions to be executed first in the current period, such as maintaining a certain electrolysis power, controlling the release rhythm of hydrogen storage, and supplementing some power generation to ensure uninterrupted oxygen supply to the hospital. After this period ends, the system compares the actual oxygen consumption, actual photovoltaic power, and actual equipment operation results with the original scenario results to determine which predictions were off and which paths are closer to actual operation. Based on this, it adjusts the prediction input and path ranking for the next period. In this way, the system does not provide a fixed strategy all at once, but continuously adjusts it based on real-time fluctuations in hospital oxygen consumption, photovoltaic power, and electricity prices. Therefore, it is more suitable for applications like medical oxygen supply, which require continuous stability but are subject to continuous disturbances from external conditions.

[0023] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An electrolytic hydrogen production buffer system for medical oxygen supply, comprising a state construction module, a prediction module, a strategy solving module, a collaborative execution module, and an update module, characterized in that: the state construction module is configured to obtain hospital oxygen data, meteorological forecast data, photovoltaic power generation prediction data, and grid electricity price data of a historical period, and electrolysis operation data, hydrogen storage state data, and power generation operation data of a current period, perform time alignment, cross-source slicing, and feature encoding, and output a multi-source state sequence; the prediction module is configured to input the multi-source state sequence into a time series attention network, perform joint extraction of oxygen consumption variation features, photovoltaic fluctuation features, and electricity price variation features, construct a scenario tree for a future period based on the joint extraction results, and output oxygen consumption prediction results, power supply prediction results, and price prediction results corresponding to each scenario node; the strategy solving module is configured to, based on the oxygen consumption prediction results, power supply prediction results, and price prediction results corresponding to each scenario node, in combination with the electrolysis operation data, hydrogen storage state data, and power generation operation data, perform a double-layer strategy search with the continuity of oxygen supply as the priority and the operation cost as the constraint, solve hydrogen production paths, hydrogen storage paths, and power generation paths corresponding to each scenario node, and output a scenario strategy set; and the collaborative execution module is configured to perform scenario backtracking consistency checking and cross-scenario common prefix extraction on the scenario strategy set, generate a current execution strategy, perform collaborative control on an electrolysis device, a hydrogen storage unit, and a power generation unit according to the current execution strategy, and output a buffer execution result of the current period.

2. The electrolytic hydrogen production buffer system for medical oxygen supply according to claim 1, characterized in that: further comprising: the update module is configured to obtain actual oxygen consumption results, actual photovoltaic results, actual electricity price results, and actual operation results corresponding to the buffer execution result, perform deviation inversion and strategy weight updating with the scenario strategy set, and write the updating results back to the time series attention network and the double-layer strategy search, and output a rolling buffer control result of a next period.

3. The electrolytic hydrogen production buffer system for medical oxygen supply according to claim 2, characterized in that: the state construction module comprises: performing time mapping and gap filling on the hospital oxygen data, meteorological forecast data, photovoltaic power generation prediction data, grid electricity price data, electrolysis operation data, hydrogen storage state data, and power generation operation data according to a unified time scale, and outputting an aligned data set; performing adjacent difference value calculation on the hospital oxygen value sequence, photovoltaic power generation prediction value sequence, and grid electricity value sequence in the aligned data set, and determining a difference value sign change time, an electricity value change time, and a connection time of historical data and current data as slicing points, dividing time slices according to time intervals between adjacent slicing points, and outputting a multi-source segment set; performing intra-slice feature extraction, inter-slice difference calculation, and cross-source change association calculation on each time slice in the multi-source segment set, concatenating state vectors in a fixed field order, and arranging the state vectors in a time sequence, and outputting the multi-source state sequence.

4. The electrolytic hydrogen production buffer system for medical oxygen supply according to claim 3, characterized in that: the prediction module comprises: ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ The multi-source state sequence is divided into continuous observation windows according to time order and input into a temporal attention network with position encoding and causal mask. The intra-head attention score is calculated for the hospital oxygen consumption field, photovoltaic power generation field and electricity price field in each observation window, and normalized weighted summation is performed and cross-attention fusion is performed with the remaining fields. The output is the oxygen consumption feature sequence, photovoltaic feature sequence, electricity price feature sequence and shared state sequence. Based on the oxygen consumption feature sequence, photovoltaic feature sequence, electricity price feature sequence and shared state sequence, a joint transition tensor for each future time is constructed. For each future time, the oxygen consumption state distribution, power supply state distribution and price state distribution are solved respectively. The three types of state distributions at the same future time are combined into candidate scenario nodes according to the time correspondence relationship. The set of candidate scenario nodes and the joint probability value of each candidate scenario node are output.

5. The electrolytic hydrogen production buffer system for medical oxygen supply according to claim 4, characterized in that: The prediction module also includes: For each candidate scene node in the candidate scene node set, calculate the node cost value, which consists of the logarithm of conditional probabilities, the sum of squared state transition differences, and the sum of squared field reconstruction differences. Sort the candidate scene nodes under the same parent node according to the node cost value, retain the first node in the sort, and write back the field differences of the remaining candidate scene nodes to the shared state sequence. Repeat the joint transition tensor construction and node cost value calculation on the shared state sequence after writing back, until the first node in the sorting is consistent in two consecutive rounds and the field difference vector is consistent. Output the corrected scene node set. In the order of future time, each correction scenario node in the correction scenario node set is connected to its corresponding parent node in the previous time step. The joint probability value of each connection is multiplied layer by layer, and the multiplication result of the nodes in the same layer is normalized to generate a scenario tree. The field values ​​corresponding to the oxygen consumption state distribution, power supply state distribution and price state distribution in each scenario node are read, and the oxygen consumption prediction result, power supply prediction result and price prediction result corresponding to each scenario node are output.

6. The electrolytic hydrogen production buffer system for medical oxygen supply according to claim 5, characterized in that: The strategy solving module includes: The oxygen consumption prediction results, power supply prediction results, and price prediction results corresponding to each scenario node are aligned with the electrolysis operation data, hydrogen storage status data, and power generation operation data according to time. A joint state point set consisting of electrolysis power, hydrogen storage flow, and power generation is constructed. The joint state point set is then connected according to the conservation of power change, hydrogen storage balance, and oxygen and power supply balance between adjacent time points. The candidate path graph corresponding to each scenario node is then output. For each path in each candidate path graph, calculate the continuity cost consisting of the sum of squares of the oxygen supply gap, the sum of squares of the hydrogen storage overrun, and the sum of squares of the state jump variables at each time point. Sort and retain each path in ascending order of continuity cost. Write the path with the highest continuity cost into the safe path set. Then, perform consistency checks on the remaining paths with the first action of the safe path set. If the consistency check fails, delete the corresponding path. Output the safe path set corresponding to each scenario node.

7. The electrolytic hydrogen production buffer system for medical oxygen supply according to claim 6, characterized in that: The strategy solving module also includes: For each path in each set of safe paths, calculate the operating cost, which is the sum of the electricity purchase cost, hydrogen production electricity consumption cost, power generation consumption cost and state switching cost at each time. Then, perform a two-level sorting of the operating cost and the continuity cost of the corresponding path in the order of continuity cost first and operating cost second, and output the first-ranked path of each scenario node as the node path result. Read the node path results corresponding to each scenario node, perform cross-scenario statistics on hydrogen production, hydrogen storage, and power generation actions at the same time, extract the action combination with the most frequent occurrence as the common action at that time, and write the common action back to the first segment of the candidate path graph corresponding to each scenario node. Repeat the continuous cost calculation, operation cost calculation, and double-level sorting for each candidate path graph after writing it back until the common action is consistent in two consecutive rounds. Output the hydrogen production path, hydrogen storage path, and power generation path corresponding to each scenario node to form a scenario strategy set.

8. The electrolytic hydrogen production buffer system for medical oxygen supply according to claim 7, characterized in that: The collaborative execution module includes: The scenario paths in the scenario strategy set are expanded into action sequences in chronological order. The differences in hydrogen production, hydrogen storage, and power generation actions of each scenario path at each time are calculated. A scenario divergence matrix is ​​generated at each time. The scenario divergence matrix is ​​associated with the scenario probability value and oxygen supply gap value of the corresponding scenario node and written into the backtracking verification table. The scenario backtracking dataset is output. Based on the scenario backtracking dataset, a prefix consistency search is performed at each time point. The row sum of the scenario divergence matrix, the weighted divergence sum of the scenario probability values, and the weighted risk sum of the oxygen supply gap values ​​are calculated cumulatively at each time point. The continuous action segments with a cumulative result of zero are identified as consistent prefix segments, and the first action time with a non-zero cumulative result is identified as the conflict time. The consistent prefix segments and conflict times are output.

9. The electrolytic hydrogen production buffer system for medical oxygen supply according to claim 8, characterized in that: The collaborative execution module also includes: At the moment of conflict, calculate the joint cost value of each candidate action group corresponding to each scenario path, which consists of oxygen supply gap, hydrogen storage surplus deviation, electrolysis switching amount, and power generation switching amount. Sort the candidates according to the joint cost value and select the first candidate action group to write to the first conflict position of the current execution strategy. Write the first conflict position of the current execution strategy back to the corresponding time in the scenario backtracking dataset. Recalculate the hydrogen production action difference, hydrogen storage action difference, and power generation action difference for each scenario path after the moment of conflict. Repeat the prefix consistency search and joint cost value calculation until all action times in the current time period are written. Output the current execution strategy. Based on the hydrogen production, hydrogen storage, and power generation actions in the current execution strategy, electrolysis control commands, hydrogen storage control commands, and power generation control commands are generated respectively. These commands are then simultaneously sent to the electrolysis unit, hydrogen storage unit, and power generation unit. The corresponding electrolysis operation results, hydrogen storage status results, and power generation operation results are read, and the buffer execution results for the current time period are output.

10. An electrolytic hydrogen production buffer system for medical oxygen supply according to claim 9, characterized in that: The update module includes: Align the actual oxygen consumption results, actual photovoltaic results, actual electricity price results, and actual operation results corresponding to the buffer execution results with the scenario oxygen consumption results, scenario power supply results, scenario price results, and scenario operation results in the scenario strategy set according to time. Calculate the oxygen consumption deviation value, photovoltaic deviation value, electricity price deviation value, and operation deviation value respectively. Merge each deviation value according to the scenario path to form a scenario deviation sequence and output the deviation inversion result. Based on the deviation inversion results, the cumulative deviation value and path consistency value are calculated for each scenario path in the scenario policy set. The cumulative deviation value and path consistency value are written into the policy weights of each scenario path. The oxygen consumption deviation sequence, photovoltaic deviation sequence and electricity price deviation sequence in the deviation inversion results are written back to the temporal attention network. The updated policy weights are written back to the two-layer policy search, and the updated results are output. Based on the updated results, scenario prediction, policy search, and cooperative control are re-executed for the multi-source state sequence in the next time period, and the rolling buffer control results for the next time period are output.

Images