A cloud-edge dual-strategy fitting and deviation backfill control method for an off-grid light storage firewood system

Abstract

The application discloses a cloud-edge double-strategy fitting and deviation back-supply control method of an off-grid light storage and diesel system, and aims at solving the problems that the cloud strategy is easy to fail, power impact is caused by strategy hard switching, and short-time control gradually deviates from long-term optimization target when communication is abnormal. The application adopts a double-end independent strategy architecture, that is, a long-time scale optimization strategy with a safety boundary is generated at the cloud end, and a short-time scale control strategy is independently generated at the edge end. The edge end dynamically allocates the double-strategy weight according to the communication reliability and the net load mutation degree, and solves the fitting control sequence in the real-time feasible region. The long-term strategy deviation is recursively accumulated, and a limited back-supply component is generated to gradually reduce the deviation when the back-supply condition is met. The cloud safety boundary constraint is reserved when the communication is interrupted, and the gradual strategy fusion is executed after the communication is restored. The method can reduce the dependence of the system on the communication link, and effectively improve the operation stability and power supply reliability of the off-grid light storage and diesel system.

Description

Technical Field

[0001] This invention relates to a cloud-edge dual-strategy fitting and deviation compensation control method for an off-grid photovoltaic-storage-diesel system, belonging to the field of off-grid microgrid energy management technology. Background Technology

[0002] Off-grid photovoltaic-storage-diesel systems consist of photovoltaic power generation units, energy storage units, and diesel generator sets, and are widely deployed in oil fields, mining areas, overseas engineering camps, disaster emergency sites, and remote villages. These scenarios are characterized by large fluctuations in renewable energy output, frequent load changes, uninterrupted power supply, and unstable communication links. The system needs to continuously maintain power balance and equipment safety under conditions of unreliable communication quality.

[0003] There are two main types of control methods for existing off-grid photovoltaic-storage-diesel systems. The first is cloud-based energy management: the cloud generates optimized scheduling results based on global data and a longer forecast time domain, and issues them in the form of discrete power commands, which are then forwarded or executed by the edge. The second is multi-timescale scheduling: long-term plans (such as day-ahead plans) are modified sequentially by short-term plans (such as intraday plans and real-time plans). Furthermore, in response to communication anomalies, existing solutions typically retain the last received command after a communication interruption, or switch control to local fixed rules; other solutions assign weights to remote conditions based on communication quality, or allow autonomous operation by the edge during network outages.

[0004] However, the applicant found that the existing methods described above have at least the following technical problems:

[0005] First, the edge device degenerates into a simple instruction executor. When communication experiences delays, jitter, packet loss, or interruptions, the policies obtained by the edge device are incomplete, out of sync, or expired, making them impossible to execute directly. Furthermore, the load, photovoltaic, and energy storage state of charge predictions used to generate cloud policies deviate from the real-time on-site conditions, leading to hard switching between cloud policies and local fixed rules, causing power command jumps in energy storage units or diesel generator sets.

[0006] Second, short-term deviations lack cross-cycle memory. Whether it is cloud-based energy management or sequential multi-timescale scheduling, deviations of short-cycle control from long-cycle plans will be forgotten in the next control cycle. When sudden changes in source load and communication instability occur simultaneously, the continuous local optimum of short-term control will gradually erode the long-term energy, state of charge, reserve capacity, and equipment start-up and shutdown intentions in the cloud strategy, resulting in the loss of long-term optimization goals.

[0007] Third, after the network is disconnected, it only operates according to fixed rules, making it difficult to retain the global optimization intentions in the cloud strategy. Moreover, after communication is restored, the old and new strategies are directly switched, which can easily cause frequent start-stop of diesel generator sets, impact charging and discharging of energy storage units, or sudden changes in bus power.

[0008] Therefore, how to align and constrain the long-term cloud strategy with the short-term edge strategy generated independently at the edge within the same control period, and how to ensure that the actual strategy deviation during mutations or communication interruptions can be accumulated and can be replenished in a limited manner after the safety conditions are restored, are technical problems that urgently need to be solved in this field. Summary of the Invention

[0009] This invention provides a cloud-edge dual-strategy fitting and deviation compensation control method for an off-grid photovoltaic-storage-diesel system to address the problems existing in the prior art.

[0010] The technical solutions adopted in this invention are as follows:

[0011] A cloud-edge dual-strategy fitting and deviation compensation control method for an off-grid photovoltaic-storage-diesel system, the off-grid photovoltaic-storage-diesel system including a cloud terminal, an edge terminal, a photovoltaic power generation unit, an energy storage unit, and a diesel generator set, comprising the following steps:

[0012] S1: The cloud generates a long-term cloud strategy based on the first prediction time domain and sends the long-term cloud strategy to the edge terminal; the long-term cloud strategy includes reference power sequences of photovoltaic power generation units, energy storage units and diesel generator sets at multiple future time nodes; the edge terminal independently generates edge short-term strategies for photovoltaic power generation units, energy storage units and diesel generator sets based on the second prediction time domain and the actual measured state on site, where the first prediction time domain is longer than the second prediction time domain;

[0013] S2: The edge maps the long-term cloud strategy to the timeline of the short-term edge strategy to obtain the cloud-aligned strategy.

[0014] S3: Calculate communication reliability and net load mutation rate at the edge, and determine cloud policy weight and edge policy weight based on communication reliability and net load mutation rate; wherein, cloud policy weight increases with the increase of communication reliability, and edge policy weight increases with the increase of net load mutation rate;

[0015] S4: At the edge, the alignment strategy with the cloud and the edge short-timescale strategy are used as parallel tracking targets. A fitting optimization problem is solved within the real-time feasible region to obtain a fitted control sequence. The objective function of the fitting optimization problem includes a first deviation term weighted by the cloud strategy weights and a second deviation term weighted by the edge strategy weights. The first deviation term is the deviation between the fitted control sequence and the alignment strategy with the cloud, and the second deviation term is the deviation between the fitted control sequence and the edge short-timescale strategy. The real-time feasible region includes power balance constraints, energy storage state of charge constraints, equipment power boundary constraints, ramping constraints, and diesel generator start-stop constraints.

[0016] S5: The edge terminal sends the first control quantity in the fitted control sequence to the photovoltaic power generation unit, energy storage unit, and diesel generator set for execution, and collects the actual received value of the equipment. The actual received value of the equipment is the power value actually executed after the first control quantity is confirmed by the equipment controllers of the photovoltaic power generation unit, energy storage unit, and diesel generator set. The edge terminal uses the actual received value of the equipment as the starting control quantity for solving the fitting optimization problem in the next control cycle.

[0017] S6: The edge device recursively updates the long-term policy deviation state based on the difference between the actual received value of the device and the reference power value of the cloud policy at the same time. When the communication reliability is higher than the compensation reliability threshold, the net load mutation degree is lower than the compensation mutation threshold, and the device safety margin meets the preset compensation conditions, the edge device generates a compensation component based on the long-term policy deviation state and incorporates the compensation component into the solution of the fitting optimization problem, gradually reducing the long-term policy deviation state under the constraints of the real-time feasible region.

[0018] Furthermore, the cloud-based long-term strategy also includes an energy storage state of charge corridor, a backup boundary, and a device allowable boundary; when communication between the cloud and the edge is interrupted, the edge retains the energy storage state of charge corridor, the backup boundary, and the device allowable boundary as constraints of the real-time feasible domain.

[0019] Furthermore, the communication reliability is determined based on communication quality, policy timeliness, and state consistency;

[0020] The communication quality is determined based on the average latency, packet loss rate, latency jitter, and the number of consecutive times a valid policy packet is not received within the sliding time window;

[0021] The timeliness of the strategy is determined based on the time difference between the current moment and the generation time of the long-term strategy in the cloud.

[0022] The state consistency is determined based on the deviation between the predicted state generated by the long-term cloud-based strategy and the actual measured state on-site.

[0023] Furthermore, the cloud policy weights and edge policy weights are determined through normalization calculations with positive zero-prevention parameters:

[0024] , ,

[0025] in, The cloud strategy weight is the weight for the k-th control cycle; The edge policy weight for the kth control cycle; This is the baseline value for cloud-based policies; the baseline value for cloud-based policies increases with increasing communication reliability and decreases with increasing net load abruptness. This is the baseline value for the edge strategy; the baseline value for the edge strategy increases with the increase of the net load abrupt change. The parameter is a positive number to prevent zeroing.

[0026] Furthermore, the objective function of the fitting optimization problem also includes a control quantity change penalty term, which is used to penalize the control quantity change between adjacent control cycles in the fitted control sequence.

[0027] Furthermore, the recursive update of the long-term strategy deviation state includes: multiplying the long-term strategy deviation state of the previous control cycle by a forgetting factor, superimposing the deviation increment obtained by time integration of the difference between the actual received value of the device and the reference power value of the cloud strategy aligned at the same time, and limiting the superposition result within the preset allowable long-term strategy deviation state boundary.

[0028] Furthermore, the objective function of the fitting optimization problem also includes a penalty term for long-term policy deviation states.

[0029] Furthermore, after communication is restored, the edge device performs version verification and validity period verification on the received new cloud-based long-term scale policy, replaces the prediction starting point of the new cloud-based long-term scale policy with the actual on-site test status, and limits the increment of the cloud policy weight in each control cycle.

[0030] Furthermore, the edge terminal progressively fuses the continuous power variables in the reference power sequence of the new cloud long-term scale strategy according to the fusion coefficient, which is increased periodically by a preset increment until it equals 1;

[0031] The start-stop state variables of the diesel generator set do not participate in the gradual fusion, and are switched only when the minimum start-up time, minimum shutdown time, and safety interlock conditions are met.

[0032] Furthermore, the replenishment component is generated according to the following formula:

[0033] , ,

[0034] in, The compensation component is the component used in the k-th control cycle. This represents the long-term policy deviation state during the k-th control cycle. To compensate for the gain, To compensate for the gating coefficient, The communication reliability for the k-th control cycle, The net load mutation rate is given in the k-th control period. This refers to the equipment safety margin index for the k-th control cycle. The saturation function is defined to limit the backfill component within a preset change boundary, where ΔU is the preset change boundary.

[0035] Furthermore, when the compensation component is incorporated into the solution of the fitting optimization problem, the net load mutation component in the compensation component is preferentially allocated to the energy storage unit and the photovoltaic power generation unit. The cross-cycle energy deviation component in the compensation component is compensated in multiple control cycles within the range allowed by the energy storage state of charge constraint and through the feasible start-stop window of the diesel generator set.

[0036] Furthermore, in S3, the edge end also adjusts the cloud policy weight according to the modulus of the long-term policy deviation state. The cloud policy weight increases as the modulus of the long-term policy deviation state increases. When the net load mutation degree is higher than the compensation mutation threshold, the adjustment of the cloud policy weight according to the modulus of the long-term policy deviation state is suspended, and the cloud policy weight is determined according to the communication reliability and the net load mutation degree.

[0037] Furthermore, when the preset replenishment conditions are not met and the long-term strategy deviation is in the accumulation period, the edge end identifies the future low-cost replenishment window based on the local short-term source load prediction, and in multiple control cycles before the arrival of the low-cost replenishment window, it gradually pre-biases the state of charge operation point of the energy storage unit in the direction that reduces the long-term strategy deviation within the real-time feasible domain.

[0038] Furthermore, the low-cost replenishment window is the period of surplus photovoltaic power, the period of low load, or the economical operating period of diesel generator sets.

[0039] The present invention has the following beneficial effects:

[0040] (1) The cloud and edge independently generate long-term cloud strategies and short-term edge strategies based on different prediction time domains, and enter the fitting optimization problem in parallel in each control cycle. When communication is normal, the edge does not degenerate into a simple actuator; when communication is abnormal, the edge still has independently generated complete short-term edge strategies available for execution, which solves the problem that the cloud strategy is incomplete or expired and cannot be executed when communication is abnormal.

[0041] (2) The cloud strategy weight and edge strategy weight are continuously adjusted with the communication reliability and net load mutation degree. The final control quantity is solved in the real-time feasible domain, rather than the linear superposition of the two strategies or the hard switching between cloud strategy and local rules. This avoids power command jumps while ensuring power balance and equipment safety.

[0042] (3) The long-term strategy deviation state is updated recursively based on the difference between the actual received value of the equipment and the reference power value of the long-term strategy in the cloud. The actual strategy deviation during abrupt changes or communication interruptions is accumulated across cycles rather than forgotten. When the communication reliability, net load mutation degree and equipment safety margin indicators meet the conditions, the backfill component is included in the solution of the fitting optimization problem. Under the constraints of the real-time feasible domain, the long-term strategy deviation state is gradually reduced so that the long-term energy, state of charge, reserve capacity and equipment start-up and shutdown intentions of the cloud strategy are not lost due to short-term control.

[0043] (4) When communication is interrupted, the energy storage state of charge corridor, reserve boundary and equipment allowable boundary in the long-term cloud strategy are retained as constraints of the real-time feasible domain. During the network outage, the state of charge, reserve capacity and equipment safety constraints are still satisfied, avoiding the loss of long-term optimization intentions caused by fixed rule operation.

[0044] (5) After communication is restored, the new cloud-based long-term strategy is verified and the prediction starting point is replaced by the actual measured state on site. The continuous power variables are gradually integrated and the start-stop state variables are switched only when the preset conditions are met, which avoids the frequent start-stop of diesel generator sets and the power impact of the bus caused by the direct switching of the old and new strategies after the network is restored.

[0045] (6) The first control quantity of the fitted control sequence is actually sent to the photovoltaic power generation unit, energy storage unit and diesel generator set for execution, and written back with the actual received value of the equipment to form a verifiable physical control closed loop. Attached Figure Description

[0046] Figure 1 This is the overall architecture diagram of the cloud-edge-unified execution controller of this invention;

[0047] Figure 2 This is a schematic diagram of the constraint policy package data structure;

[0048] Figure 3 This is a flowchart of the method of the present invention;

[0049] Figure 4 This is a flowchart of the joint calculation of communication trustworthiness based on communication quality, policy timeliness, and state consistency;

[0050] Figure 5 This is a schematic diagram illustrating the mapping of cloud-based long-scale strategies to the edge timeline and the handling of missing nodes.

[0051] Figure 6 It is a state machine diagram of normal, degradation, autonomous, network integration and security protection;

[0052] Figure 7 This is a flowchart of policy completion and local autonomy during communication interruption;

[0053] Figure 8 This is a flowchart of state alignment, progressive fusion, and feasible domain re-verification after communication recovery;

[0054] Figure 9 It is a diagram showing the command and feedback relationships between the photovoltaic inverter, the energy storage PCS, and the diesel generator set controller;

[0055] Figure 10 It is a graph showing the cloud-based long-term strategy, hold value, fitting control, and edge short-term correction.

[0056] Figure 11 This is a diagram of a communication fault injection and control verification platform based on the Simulink / Simscape model;

[0057] Figure 12 It is a graph showing the duration of communication interruption, the weight of cloud policy and the weight of edge policy, the long-term policy deviation status and the compensation component.

[0058] Figure 13 It is a diagram showing the actual power distribution of the diesel generator set, battery energy storage system (BESS), photovoltaic system, and load.

[0059] Figure 14 This is a diagram of the in-model SOC observation results based on actual BESS power;

[0060] Figure 15 These are the official original control and cloud-edge control frequency, voltage, and BESS three-phase current response diagrams;

[0061] Figure 16 This is a comparison chart of key performance indicators under the same physical model;

[0062] Figure 17 It includes cloud strategy weights and edge strategy weights, compensation components, and power tracking diagrams for diesel generator sets.

[0063] Figure 18 This is a BESS three-phase current phase-by-phase, cycle-by-cycle RMS and peak value diagram;

[0064] Figure 19 It is an event state linkage diagram of communication interruption, policy preservation, fitting optimization, and component compensation;

[0065] Figure 20 This is a combined response diagram of BESS active power, reactive power, current cycle-by-cycle RMS and SOC energy state;

[0066] Figure 21 It is a diagram of active / reactive resource collaborative allocation and power balance residuals;

[0067] Figure 22It is a matrix diagram showing the changes in key performance indicators at different communication and load stages. Detailed Implementation

[0068] The invention will now be further described with reference to the accompanying drawings.

[0069] like Figure 1 , Figure 2 and Figure 3 This invention provides a cloud-edge dual-strategy fitting and deviation compensation control method for an off-grid photovoltaic-storage-diesel system. The off-grid photovoltaic-storage-diesel system includes a photovoltaic power generation unit, an energy storage unit, a diesel generator set, a cloud platform, an edge platform, and a unified execution controller. The photovoltaic power generation unit, energy storage unit, and diesel generator set are connected to the edge platform through the unified execution controller, and the edge platform is connected to the cloud platform through a communication link.

[0070] The system consists of three layers: a cloud-based mainframe (i.e., the cloud), an edge-based secondaryframe (i.e., the edge), and a unified execution controller. The cloud-based mainframe, deployed in a remote data center or cloud platform, is responsible for generating long-term cloud-based strategies based on global historical data, long-term source load prediction, and cross-period constraints. The prediction time domain of the cloud-based long-term strategy is denoted as H. c For example, the next 6 to 24 hours; the strategy step size is denoted as Δt. c For example, 15 to 60 minutes; the strategy update cycle is for example, 30 to 120 minutes or event-triggered.

[0071] The edge sub-brain is deployed on industrial computers, ARM edge boxes, or field controllers with real-time operating systems. It is responsible for independently generating short-term edge strategies based on actual field conditions, local short-term source load predictions, and real-time device boundaries. It also performs fitting and optimization between the long-term cloud strategies and the short-term edge strategies. The prediction time domain of the short-term edge strategy is denoted as H. e For example, the next 5 to 60 minutes; the strategy step size and control period are denoted as Δt. e For example, the time domain and step size are from 1 second to 5 minutes. The prediction time domain and step size satisfy... and .

[0072] The edge auxiliary computer connects to the unified execution controller via Ethernet, CAN, or serial bus. The unified execution controller, which can be an industrial PLC, real-time controller, or self-developed control board, is responsible for protocol adaptation (Modbus TCP / RTU, CAN, RS485, or vendor protocol), command limiting, safety interlocking, protocol conversion, and equipment feedback acquisition. The unified execution controller connects to the photovoltaic inverter, energy storage converter (PCS), battery management system (BMS), and diesel generator set controller through corresponding communication interfaces.

[0073] The edge computing unit sends the first control quantity of the fitted control sequence to the unified execution controller. This control quantity includes at least the active power target of the photovoltaic power generation unit, the charging and discharging power target of the energy storage unit, and the power / start / stop target of the diesel generator set. The unified execution controller converts the control quantity into a protocol format recognizable by each device, performs amplitude limiting and interlock verification, and then sends it to the corresponding device. It also collects the actual received value from the device and sends it back to the edge computing unit. The actual received value refers to the power value actually executed after the first control quantity is confirmed by the device controller. If the device limits the execution of the control quantity due to temperature derating, fault protection, or power boundary limitations, the actual power after limiting is used as the actual received value.

[0074] The cloud-based master brain outputs long-term cloud-based policies, while the edge-based slave brain simultaneously outputs independent short-term edge-based policies. The edge-based slave brain first interpolates or maintains the long-term cloud-based policies onto the edge time axis, then feeds both policies, the current device state, communication reliability, and net load abruptness into the fitting optimizer to obtain a unique executable device control sequence. The final control is not an unconstrained linear superposition of the two policies, but rather a fitting optimization result with physical constraints, continuity terms, and state corridor terms.

[0075] ,

[0076] in, This represents the optimal fitted control input for the k-th control cycle. Let be the real-time feasible region for the k-th control cycle. Let be the objective function of the fitting optimization problem for the k-th control cycle.

[0077] The cloud-based main brain receives inputs including global historical data, long-term source load predictions, device rated parameters, and edge backhaul summaries, and outputs a policy package. The edge secondary brain receives inputs including on-site measured conditions, local short-term source load predictions, a policy package, and link statistics. It first independently outputs short-term edge policies, and then outputs the first control variable after fitting optimization and feasible region verification.

[0078] The unified execution controller returns the device's acceptance, limiting, rejection, or fault status; the edge auxiliary brain updates the starting control quantity and long-term strategy deviation status of the next control cycle with the actual accepted value of the device, so that unrealized cloud energy, energy storage state of charge (SOC), or backup intentions are not lost due to the rapid control cycle rolling.

[0079] The control priorities, from highest to lowest, are: hardware protection and personal safety, emergency support for bus voltage and frequency, equipment interlocking and real-time feasible domain, online fitting of cloud-edge dual strategies, long-term strategy deviation state limitation compensation, and economic optimization.

[0080] Accumulated deviations are allowed when high-priority conditions are triggered, and forced remediation is not required during abrupt changes; long-term policy deviations are gradually reduced only when communication reliability, net load abrupt changes, and equipment safety margins meet the conditions.

[0081] The cloud-based master brain encapsulates the generated long-term policies into policy packages and distributes them to the edge-based secondary brains. The policy package is denoted as:

[0082] ,

[0083] Where v is the policy version number; For strategy generation time; The timestamp sequence of the policy nodes; Reference power and diesel generator status for each node's photovoltaic, energy storage, and diesel generator sets; To optimize the corresponding predicted state sequence for the cloud, it should include at least the energy storage state of charge (SOC), load power, photovoltaic available power, and diesel generator set status; , These are the allowable lower and upper bounds for each control variable; Constraint parameters such as energy storage state of charge corridor, reserve boundary, ramp rate, minimum start-up time and minimum downtime; The set of allowed operating modes; The confidence level of the strategy node or prediction result; The effective time interval of the strategy; For integrity verification or digital signature.

[0084] After receiving the policy package, the edge sub-brain performs the following verifications in sequence: whether the data length is consistent with the number of time nodes; whether each reference quantity is within its corresponding upper and lower boundaries; whether there is an inversion of the upper and lower limits of the energy storage state of charge corridor; whether the diesel generator set state sequence meets the start-stop constraints; whether the version number is later than the cached version; whether the timestamp is monotonic; and whether the validity period covers the current time. If a field is missing but does not affect the safety boundary, it can be filled in using the corresponding field of the previous compatible version or the edge default value; if key fields involving power boundaries, SOC safety boundaries, or integrity verification are missing, the policy package will not be used as a strong tracking target.

[0085] Each time point in the policy package can carry a node-level confidence score and a valid bit. The peripheral subbrain constructs a validity marker for the i-th node:

[0086] ,

[0087] in, This is an indicator function; it takes the value 1 when the condition within the parentheses is met, and 0 otherwise. The minimum confidence threshold; Mark the validity of the i-th node; and These are the allowed lower and upper bounds for the i-th node, respectively. When all three criteria are met, the node is added to the set of valid nodes; otherwise, the node is only recorded as historical data and not as a parallel tracking target.

[0088] like Figure 4 The limbic subbrain calculates communication quality, including average latency, within a sliding time window. Packet loss rate latency jitter Number of consecutive times no valid policy packets were received And optional effective bandwidth or signal strength indicators.

[0089] Communication quality is calculated using the following formula:

[0090] ,

[0091] in, These are the component weight coefficients, and they satisfy... ; This is the normalized baseline value.

[0092] The timeliness of a strategy is calculated using the following formula:

[0093] ,

[0094] in, For the current moment, For strategy generation time, This is the time constant for policy effectiveness decay. When the current time exceeds the end of the policy's effective time interval... At the same time, strategies should not be used as the sole basis for control.

[0095] State consistency compares the predicted state when the policy is generated with the current measured state:

[0096] ,

[0097] in, This is the current measured state vector. This refers to the predicted state at the corresponding time in the cloud-based policy package. This is the state weight matrix. The state vector must include at least the energy storage SOC, available PV power, load power, diesel generator state, and current output power. Variables that will lead to power outages, SOC exceeding limits, or diesel generator state conflicts are assigned higher weights.

[0098] communication credibility Calculate using the following formula:

[0099] ,

[0100] in, , , These are the component weight coefficients, and they satisfy... ; It is a saturation function limited to the interval [0,1].

[0101] To suppress sudden changes in credibility caused by a single anomalous data packet, a first-order smoothing is applied to the original communication credibility:

[0102] ,

[0103] in, The original communication reliability without smoothing in the k-th control cycle. The smoothed communication reliability is represented by μ, where μ is the smoothing coefficient.

[0104] Use a lower threshold to enter a communication degradation state Use a higher threshold to exit the degenerate state. And satisfy The state switch is only executed when the threshold condition is maintained continuously for a preset time, thereby avoiding frequent changes in control mode by the edge sub-brain when the link quality fluctuates near the threshold.

[0105] Cloud-based long-scale strategies are denoted as It includes at least multiple future time points, such as reference power sequences, predicted state sequences, energy storage state-of-charge corridors, reserve boundaries, equipment allowable boundaries, version numbers, and effective times for photovoltaic power generation units, energy storage units, and diesel generator sets. The edge computing system, in the k-th control cycle, bases its decisions on the actual measured state at the field. Local short-time source load prediction and real-time device boundary independent solution edge short-timescale strategy The edge short-timescale strategy continues to generate even when cloud communication is normal, rather than only being activated after communication is interrupted:

[0106] ,

[0107] in, Optimize the objective function for short-timescale edge strategies. This is for local short-term source-load forecasting. The edge short-timescale strategy prioritizes reflecting the current power balance, frequency / voltage support, SOC safety margin, physical status of diesel generator sets, and equipment derating. Its prediction time domain and step size are smaller than those of the cloud strategy.

[0108] like Figure 5The edge sub-brain trims the cloud policy according to the policy's effective time interval to the current moment and maps it to the edge time axis using zero-order preservation, piecewise linear, spline, or Hermite interpolation methods to obtain an aligned cloud policy. The mapping only completes the time alignment and does not directly generate the first control variable.

[0109] The alignment operations include: trimming cloud policies to the current time based on their version validity period and interpolating or maintaining them to the edge time axis; unifying cloud and edge policies to the same device variable order, power sign, and discrete state encoding; using the current measured SOC, diesel generator set status, and available device power as the initial state for fitting; reducing the validity of missing or expired cloud power nodes, but retaining the still valid energy storage state of charge corridor and safety boundary.

[0110] Cloud timeline mapping is represented by the following formula:

[0111] ,

[0112] in, For interpolation or preservation operators, Predict the number of nodes in the time domain for edge strategies. The mapping only performs time alignment and does not directly generate the final device control variables.

[0113] limbic brain calculation of net load mutation degree The net load mutation rate is calculated by the following formula as an example:

[0114] ,

[0115] in, The net load for the k-th control cycle, This refers to the system's rated power or reference load power. The parameter is a positive zero-prevention parameter. The net load mutation rate is normalized to the [0,1] interval and the absolute value is taken to ensure that it is positive when the load suddenly increases or decreases.

[0116] Cloud strategy benchmark and edge strategy benchmark Calculate using the following formula:

[0117] ,

[0118] ,

[0119] in, For communication quality, For the timeliness of the strategy, For state consistency, For local short-term forecast error estimation, , , This represents the weighting coefficient. The baseline value for cloud-based policies increases with increasing communication reliability and decreases with increasing net load abruptness; the baseline value for edge-based policies increases with increasing net load abruptness.

[0120] Cloud-based strategy weights and edge strategy weights Determined through normalization calculation with a positive zero-prevention parameter ε:

[0121] , ,

[0122] in, The cloud policy weight for the k-th control cycle. The edge policy weights for the k-th control cycle are... , and .

[0123] The decision variables for the fitting optimization problem include:

[0124] ,

[0125] in, The active power of the photovoltaic power generation unit. The charging and discharging power of the energy storage unit (discharging is positive, charging is negative). For diesel generator set power, This refers to the shutdown / operation status of the diesel generator set. Decision variables must include at least photovoltaic active power, energy storage charging / discharging power, diesel generator set power and its start / stop status; reactive power, grid connection / connection mode, and interruptible loads can be added. A superscript indicates that the variable is a decision variable for an online fitting optimization problem, that is, the optimal reference value to be solved by the optimizer in the current control cycle; a variable without a superscript indicates the value of the physical quantity in the actual constraint equation, state transition or known conditions.

[0126] The power balance constraint is:

[0127] ,

[0128] in, For unmet loads (power outages), it is located on the generation side to make up for the power balance; To allow for power imbalance relaxation that exists during the optimization process but is penalized by high weights.

[0129] Energy storage state constraints are:

[0130] ,

[0131] And satisfy:

[0132] ,

[0133] in, For charging efficiency, For discharge efficiency, For the rated energy of energy storage, To obtain the positive part operator (i.e. The upper and lower limits of SOC can be adopted from the energy storage state of charge corridor issued by the cloud-based main brain, and can be appropriately tightened by the edge auxiliary brain when the prediction error increases.

[0134] Diesel generator set constraints include:

[0135] ,

[0136] ,

[0137] Simultaneously set constraints for minimum startup time, minimum shutdown time, warm-up time, and maximum number of start / stop cycles.

[0138] Calculate the upward and downward adjustment of reserve demand based on the load and photovoltaic forecast error range:

[0139] ,

[0140] ,

[0141] It also requires that the remaining charging and discharging capacity of energy storage and the adjustment capacity of available diesel generator sets cover the corresponding backup needs.

[0142] in, This refers to the minimum permissible output power of the diesel generator set; This refers to the maximum permissible output power of the diesel generator set. This represents the maximum downhill gradient of the diesel generator set. This represents the maximum gradeability of the diesel generator set. The edge strategy step size.

[0143] For the first Increased reserve requirements for each control cycle; For the first Reduced reserve requirements for each control cycle; This is the reserve demand coefficient; This represents the upper limit of the predicted load power for the k-th control cycle. For the first Lower limit of load power prediction for each control cycle; This is the lower limit for predicting available photovoltaic power in the k-th control cycle; For the first Upper limit of photovoltaic available power prediction for each control cycle; For the first Net load forecast value for each control cycle.

[0144] The objective function of the fitting optimization problem simultaneously penalizes deviations of the fitted control sequence from the aligned long-term cloud-based strategy, deviations from the edge short-term strategy, changes in control variables, and deviations from the long-term cloud-based corridor; all terms are weighted after unifying dimensions or normalization. This process differs from directly weighting and adding two instructions because the decision variables must simultaneously satisfy power system dynamics and equipment constraints.

[0145] The objective function of the fitting optimization problem has the following structural form:

[0146] ,

[0147] in, This is a cloud-based strategy deviation item. For edge strategy deviation term, Penalty for changes in the control quantity, For the state corridor deviation term, This is a penalty term for long-term strategy deviation. For the first The total value of the objective function for fitting the optimization problem over a control cycle. For the first The desired fit control sequence to be optimized over a control cycle. For the first A time-aligned cloud-based long-term strategy reference sequence for each control cycle. For the first Each control cycle's limbic secondary brain independently generates a short-term limbic timescale policy reference sequence. For the first A sequence of system states to be optimized in a control cycle. For the first The state corridor target sequence is distributed from the cloud in each control cycle. For the first The long-term policy deviation state sequence to be optimized in each control cycle.

[0148] Expanding the above terms to the prediction time domain and summing them node by node, we get:

[0149] And it satisfies the state transition equation and constraints:

[0150]

[0151] in, The cloud-based policy weight increases with improved communication trust and state consistency. The edge strategy weights increase with the increase in net load abruptness and short-term forecast error; , , , , Let be the positive semidefinite weight matrices for cloud policy deviation, edge policy deviation, control change, state corridor deviation, and long-term policy deviation states, respectively; U is the fitted control sequence decision variable over the entire prediction time domain; X is the system state sequence decision variable over the entire prediction time domain; Z is the long-term policy deviation state sequence decision variable over the entire prediction time domain; and j is the node index in the prediction time domain, taking values... ; For the first The fitted control quantity for each control cycle; For the first Aligned cloud strategy reference control quantity after interpolation or maintenance for each control cycle; For the first The edge short-timescale policy reference control quantity for each control cycle; For the first The first control cycle and the first Changes in control quantity between control cycles; For the first The system state vector for each control cycle; For the first The target value of the status corridor is sent down from the cloud in each control cycle; For the first The long-term policy deviation state vector for each control cycle. This is the system state transition function, which describes the evolution of the system state with respect to control and disturbance variables. It is an equality constraint function used to describe equality constraints that must be strictly satisfied, such as power balance; These are inequality constraint functions used to describe inequality constraints such as device power boundary, SOC boundary, and ramp boundary. Let J be the external disturbance vector of the j-th node, which includes at least the load power demand and the available photovoltaic power, and take the local short-term source load forecast. The value taken at the corresponding node.

[0152] After this round of execution, the initial value for the next cycle is recursively corrected based on the difference between the actual value received by the device in this round and the cloud-aligned strategy at the same time.

[0153] The limbic subbrain executes only the first control variable of the fitted control sequence each time. After being issued by the unified execution controller and confirmed by the device controller, it enters the next cycle based on the actual value received by the device and the new measured state.

[0154] ,

[0155] in, For operators that project onto the real-time feasible region, This is the device confirmation function.

[0156] To avoid continuous and rapid optimization that leads to the forgetting of long-term cloud energy, SOC, backup, and start / stop intentions, the edge sub-brain maintains the long-term strategy bias state. (Also known as the policy deviation ledger). This state does not use optimizer-calculated values, but is updated based on the difference between the actual value received by the device and the reference power value aligned with the cloud policy at the same point in time:

[0157] ,

[0158] in, The forgetting factor has a value range of (0,1]. This is a deviation mapping matrix used to convert the power deviation of each device into energy, SOC, standby, or start / stop intention deviation; For the first The actual value received by the equipment in each control cycle; For the first The reference power value for aligning the cloud strategy with each control cycle; For edge strategy step size; This is a saturation function designed to limit long-term policy deviation states within a preset allowed long-term policy deviation state boundary.

[0159] Deviation mapping matrix Constructed as an example of a block diagonal matrix: power to SOC deviation according to The conversion is as follows: the deviation from power to standby intention is mapped according to the adjustable margin of each device, and the deviation from power to start-stop intention is mapped according to the cost of diesel generator set state switching.

[0160] When communication is interrupted or the net load changes abruptly, short-term edge strategies can quickly take over within the real-time feasible region, while long-term strategy deviations only accumulate deviations without forcing immediate regression. When communication reliability is higher than the recovery reliability threshold, the net load change rate is lower than the recovery change rate threshold, and the device safety margin meets the preset recovery conditions, a recovery component is generated.

[0161] ,

[0162] ,

[0163] in, For the first The compensation component of each control cycle To compensate for the gain, To compensate for the gating coefficient, Smoothed communication reliability , For equipment safety margin indicators, This is a saturation function that limits the backfill component to a preset change boundary ΔU.

[0164] Equipment safety margin indicators For example, the normalized minimum values ​​of the energy storage power change rate margin, diesel generator ramping margin, start-stop time slack, and bus power change margin are taken as follows:

[0165] ,

[0166] in, This is the rated power of the energy storage. For the first The actual received value of the energy storage unit in each control cycle; For the first The power of the diesel generator set in each control cycle; No. The power of the diesel generator set in each control cycle This represents the remaining time between the current state of the diesel generator set and the next permitted switchover. Minimum start-stop time, To allow for the maximum variation in bus power, No. The change in bus power per control cycle.

[0167] The compensation component is incorporated into the solution of the fitting optimization problem, and the long-term strategy deviation state is gradually reduced under the constraints of the real-time feasible region. The net load mutation component is preferentially allocated to fast-response equipment such as energy storage units and photovoltaic power generation units; the cross-cycle energy deviation component is gradually compensated in multiple control cycles through the energy storage state of charge corridor and the feasible start-stop window of the diesel generator set, without forcing slow equipment to follow in a single-cycle hard switching manner.

[0168] The real-time feasible region of the kth control cycle can be represented as:

[0169] ,

[0170] Among them, the rated boundary Provided by the equipment nameplate and controller configuration; derating boundary Real-time updates based on temperature, fault status, and available power in the BMS / PCS; SOC boundary. Including safety upper and lower limits and dynamic corridors; diesel generator set boundaries. Including preheating, minimum start / stop time, and paralleling conditions; ramp boundary. and spare boundary Determined by equipment physical constraints and system safety requirements. The edge computing module is updated based on the latest feedback. Then, solve the fitting optimization problem to avoid using device capacity that has expired when generated in the cloud.

[0171] When the fitting optimization problem has no solution, the cloud strategy weights should be reduced in the following order: relax the unsafe state corridor, allow limited light curtailment, trigger diesel generator preheating or parallel operation, and manage non-critical loads; the hardware rated boundaries, SOC hard safety boundaries, or diesel generator safety interlocks must not be relaxed. The edge optimizer should be set with a maximum solution time. .exist If a feasible but non-optimal solution is found, that feasible solution is executed first; if a timeout occurs and no feasible solution is found, the previously verified control variable is used and the amplitude is re-limited according to the latest device boundary, while a fast safety rule is activated. This ensures that the optimization calculation does not block the underlying real-time control.

[0172] like Figure 6 and Figure 7 The system includes at least the following five states:

[0173] C0 normal cloud-edge collaboration, communication quality, policy timeliness and state consistency are all higher than normal thresholds. Long-term cloud-based policies maintain high weight, while short-term edge-based policies continuously generate and process small disturbances in the field. The two are fitted online, and the exit condition is that the credibility continues to decrease or the protection is triggered.

[0174] C1 communication degradation occurs when latency, packet loss, or jitter exceeds the degradation threshold and persists for a preset time, but there are still effective policy nodes. The cloud policy weight is reduced and the edge policy weight is increased. Time alignment is completed using the cache boundary. Energy storage and diesel generator sets are reserved in advance for backup. The exit condition is link recovery or interruption.

[0175] C2 edge autonomy: If no valid policy is received within the communication interruption judgment time, or if the policy has expired and cannot be updated, the energy storage charge state corridor, safety boundary and allowed operation mode in the most recent cloud policy are retained. The edge short time scale policy dominates the fitting optimization. If necessary, diesel generator preheating, parallel operation or non-critical load management is triggered. In this state, the adjustment of cloud policy weights based on the modulus of long-term policy deviation is suspended. The exit condition is that a valid new policy is received continuously and the verification is passed.

[0176] C3 network integration continuously receives a preset number of valid new strategy packages and completes version, validity period and feasibility verification. The prediction starting point of the new cloud strategy is replaced with the actual test status. The increment of cloud strategy weight is limited every cycle. The participation of edge short time scale strategies is maintained until the deviation of the two types of strategies continuously meets the standard. The exit condition is that the residual of strategy deviation and long-term strategy deviation status continuously meets the standard.

[0177] C4 safety protection predicts that there is no feasible solution, or that safety events such as SOC overrun risk, bus frequency / voltage overrun, and equipment failure will occur. Safety rules take precedence over cloud and edge optimization. The unified execution controller will execute energy storage rapid support, diesel generator set commissioning, photovoltaic power limiting or non-critical load disconnection according to the preset protection sequence. The exit condition is that the fault is eliminated and the status is feasible.

[0178] State switching employs a hysteresis criterion based on upper and lower thresholds and duration. C4 safety protection has preemptive priority over other states. Switching between C0 and C3 is not allowed to bypass integrity and feasibility checks. After discrete actions such as diesel generator preheating, paralleling, and shutdown are initiated, the corresponding minimum duration must be met; immediate reverse action is not permitted due to a short-term communication state bounce.

[0179] like Figure 8 After communication is restored, the edge sub-brain sequentially performs the following checks on the new strategy package: integrity, signature, version number, and validity period verification; shifts or trims the new strategy timeline to the current moment; calculates the feasibility of the new strategy based on the current SOC, diesel generator set status, and equipment power; calculates the deviation between the new strategy reference quantity and the current control quantity; and generates a fusion reference trajectory. For continuous variables such as the active / reactive power of photovoltaic power generation units, energy storage units, and diesel generator sets, gradual fusion is performed according to the fusion coefficient.

[0180] ,

[0181] in, To integrate the reference trajectory, For the current edge strategy, For the new cloud strategy, is the fusion coefficient for the k-th control cycle.

[0182] Linear fusion is only used for continuous reference variables. Discrete variables such as diesel generator start-up and shutdown, preheating, parallel operation, and grid connection / follow-the-grid modes are not linearly fused, but are switched by a state machine that meets the preheating time, minimum start-up time, minimum shutdown time, and safety interlock conditions.

[0183] The fusion progress of continuous variables is limited by the following formula:

[0184] ,

[0185] in, It is determined by the energy storage power change rate, the diesel generator set gradeability, the diesel generator set start-stop constraints, and the allowable bus power change.

[0186] Instead of directly issuing the fused reference trajectory, it is fed back into the real-time feasible domain for optimization. When the tracking error requirements are met continuously, it switches to normal cloud-edge collaborative state.

[0187] After communication is restored, the cloud timeline is not simply moved to the current moment; the cloud prediction starting point must also be replaced with the current SOC, diesel generator physical state, available photovoltaic power, and equipment derating status. Nodes missed in the cloud strategy are pruned; nodes that conflict with the discrete state on site are marked as needing recalculation or interpreted by the state machine. Continuous variables such as photovoltaic active power, energy storage charging and discharging power, and diesel generator power can be gradually transitioned using fusion coefficients. Discrete variables such as diesel generator start-up and shutdown, preheating, parallel operation, and grid / grid connection modes do not participate in linear fusion; they are switched only after the corresponding preconditions, minimum duration, and interlocking conditions are met.

[0188] Combination Figure 3 The method of the present invention includes the following steps:

[0189] S1: Data acquisition and state construction. The edge sub-brain acquires load power, photovoltaic available power, energy storage SOC, energy storage temperature, diesel generator set operating status, equipment power, bus frequency, voltage and communication link indicators, and constructs the system state vector.

[0190] S2: Cloud-based long-term strategy generation. The cloud-based mainframe generates device reference sequences, state corridors, and safety boundaries based on global data and cross-cycle constraints over a longer prediction time domain. The cloud uses point or interval predictions of load and photovoltaic power as input to establish a long-term optimization model for off-grid photovoltaic-storage-diesel systems. An example objective function is:

[0191]

[0192]

[0193] in, For the first The cost of running the cloud step-by-step at each time point; Optimize the objective function for long-term cloud computing. This refers to the startup and operation cost coefficient for diesel generator sets. For cloud strategy step size; For the first Available photovoltaic power at each time point; This is the penalty coefficient for discarded light; This is the penalty coefficient for failure to provide supplies; This is the penalty coefficient for insufficient reserves; This is the index of the time node within the cloud-based prediction time domain, with a value of [value]. ; Predict the total number of nodes in the time domain in the cloud; The cost of fuel for diesel generator sets; For the first The start-up and shutdown variables of the diesel generator set at each time point; This is the energy storage throughput or degradation cost coefficient; Insufficient reserves; For the terminal target SOC; This is the terminal SOC deviation penalty coefficient; To optimize the energy storage state of charge at the end of the time domain, that is, the end state reached after N state transitions from the initial state at the 0th time node.

[0194] Cloud-based optimization must at least meet power balance constraints:

[0195]

[0196] in, For the first The active power of a photovoltaic power generation unit at each time point; For the first The charging and discharging power of the energy storage unit at each time point; For the first The power of the diesel generator set at each time point; For the first Unmet load (power outage) at each time point; For the first Load power at each time point; For the first The system power loss at each time point.

[0197] Cloud-based optimization also satisfies constraints on available PV power, energy storage charging and discharging power, SOC, diesel generator output, ramp rate, minimum start-up and shutdown time, and up / down reserve capacity. The cloud-based mainframe outputs not only the optimal power point but also the equipment's allowable boundaries, energy storage state-of-charge corridors, and a set of operating modes, ensuring that the edge-based auxiliary brains retain long-term optimization intentions even in the event of communication anomalies. Specific prediction or optimization algorithms are not limited.

[0198] S3: Edge short-term strategies are generated independently. The edge sub-brain generates edge short-term strategies independently within a shorter prediction time domain, based on the actual measured state, local short-term source load prediction, and real-time device boundaries.

[0199] S4: Cloud strategy encapsulation, distribution, and time alignment. The long-term cloud strategy is encapsulated into a strategy package with timestamps, boundaries, validity period, and version. The edge sub-brain trims and maps it to the edge strategy timeline to obtain the aligned cloud strategy.

[0200] S5: Communication reliability and payload mutation rate calculation. The edge sub-brain calculates communication quality, policy timeliness, state consistency, payload mutation rate, and short-term prediction error to obtain cloud policy weights and edge policy weights. When the magnitude of the long-term policy deviation state increases and the device safety margin allows, and the communication reliability is higher than the degradation exit threshold, the cloud policy weight is increased, forming a negative feedback of "the greater the deviation, the stronger the regression intention". When the payload mutation rate is higher than the recovery mutation threshold, the adjustment of the cloud policy weight based on the magnitude of the long-term policy deviation state is paused, and the cloud policy weight is determined based on the communication reliability and payload mutation rate.

[0201] S6: Cloud-edge dual-strategy constrained fitting optimization, with the alignment of cloud strategy and edge short-timescale strategy as parallel tracking objectives, solves for the fitted control sequence within the real-time feasible region, taking into account long-term state corridor, short-term power mutation response, and command continuity. The initial control variable of the fitting optimization problem is fixed as the actual received value of the device in the previous cycle.

[0202] S7: First control quantity execution and device actual accepted value write-back. The edge sub-brain sends the first control quantity in the fitted control sequence to the unified execution controller, which executes it after protocol conversion, limiting, and interlocking. It collects device acceptance, limiting, rejection, or fault information, and uses the device's actual accepted value as the starting control quantity for solving the fitted optimization problem in the next control cycle.

[0203] S8: Handling communication degradation or interruption. When communication degrades, the weight of cloud policy is continuously reduced. When communication is interrupted or policy expires, strong tracking of cloud power nodes is stopped, and edge policy takes over. At the same time, long-term policy deviation status continues to be accumulated according to the actual value received by the device.

[0204] S9: Communication recovery processing. After communication is restored, the cloud prediction starting point is replaced with the actual measured status on site, and the incremental weight of the cloud strategy is limited. The long-term strategy deviation status is only subject to limited compensation when the communication credibility meets the conditions, the net load mutation degree subsides, and the equipment safety margin index meets the preset compensation conditions.

[0205] S10: Data summary backhaul and cloud master brain update. Edge sub-brains backhaul cloud long-term scale strategy, edge short-term scale strategy, fitting results, actual device acceptance values ​​and bias attribution. The cloud master brain updates the next version of cloud long-term scale strategy and boundary accordingly.

[0206] S11: Policy Bias Attribution. The edge computing layer categorizes biases into five types: prediction error, communication loss, device rejection, protection limiting, and fault derating. It records the timestamp, duration, and impact for each type of bias. The cloud only uses attributable data to update the prediction model or constraint parameters, avoiding the mislearning of device faults as normal load patterns.

[0207] The cloud-based mainframe includes modules for data processing, load and photovoltaic forecasting, long-term optimization, strategy package generation, version and integrity management, and operational deviation learning. The edge-based secondaryframe includes modules for state acquisition, communication quality evaluation, communication reliability calculation, strategy caching and time alignment, edge short-term strategy generation, dual-strategy online fitting and optimization, state machine, strategy weight migration and network alignment, and operational data summarization and retransmission. The unified execution controller includes modules for protocol adaptation (Modbus / CAN / RS485), photovoltaic, energy storage, and diesel generator command conversion, command limiting and interlocking, voltage, frequency, and power feedback, and local security protection.

[0208] like Figure 9 The command and feedback relationships between the edge sub-brain, the unified execution controller, the photovoltaic inverter, the energy storage PCS, and the diesel generator set controller are as follows: The edge sub-brain outputs the first control quantity of the fitted control sequence to the unified execution controller; after completing protocol adaptation, limiting, and interlocking, the unified execution controller sends active / reactive power targets to the photovoltaic inverter, charge / discharge power targets to the energy storage PCS, and power / start / stop targets to the diesel generator set controller; each device controller returns the actual accepted value, limiting status, rejection status, or fault status of the device to the unified execution controller; the unified execution controller transmits the actual accepted value of the device back to the edge sub-brain.

[0209] The edge processor can be deployed on an industrial computer, an ARM edge box, or a controller with an operating system; the unified execution controller can be an industrial PLC, a real-time controller, or a self-developed control board. The edge processor and the unified execution controller can communicate via Ethernet, CAN, or a serial bus; the unified execution controller connects to the photovoltaic inverter, PCS / BMS, and diesel generator set controller via Modbus TCP / RTU, CAN, RS485, or vendor protocols. This invention is independent of specific chips or protocols. Each set of device instructions includes at least a target value, allowable slope, valid time, serial number, and timeout action. The device returns an accept, reject, limited execution, or fault status. The edge processor updates the status with the actual accepted value of the device; if no confirmation is received for several consecutive cycles, the available power of the corresponding device is reduced to a safe value, and the real-time feasible domain is reconstructed.

[0210] Example 1: Online fitting of cloud-edge dual-strategy under normal communication.

[0211] The cloud-based mainframe generates a long-term cloud-based policy for the next 24 hours every 60 minutes, with node intervals of 15 minutes. The edge-based auxiliary brain independently generates a short-term edge-based policy every minute based on the actual on-site conditions and local short-term source load predictions for the next 30 minutes. The edge-based auxiliary brain maps the long-term cloud-based policy to a 1-minute time axis, performs fitting optimization once per minute, and executes the first control variable. Under normal communication conditions, the cloud-based policy has a higher weight, while the edge-based short-term policy handles small on-site disturbances; both are constrained within the real-time feasible region. This verifies that both types of policies participate in control simultaneously under normal communication conditions, and the edge-based auxiliary brain does not degenerate into a simple actuator.

[0212] Example 2: Random packet loss and high latency.

[0213] Random packet loss, sudden continuous packet loss, delays of 200 milliseconds to several seconds, and jitter are injected into the communication link. Decreased communication reliability causes a continuous decrease in the weight of cloud-based policies and a continuous increase in the weight of edge-based policies with short timescales. The edge sub-brain utilizes cached energy storage state-of-charge corridors and safety boundaries to maintain constraints, avoiding power command jumps caused by direct switching to fixed rules. Compared to maintaining the previous command or fixed rule control, this invention performs better in terms of power supply failure, SOC deviation, and the number of diesel generator start-stop cycles.

[0214] Example 3: Complete Communication Interruption Autonomy.

[0215] Communication is cut off before or after load increases or sudden drops in sunlight. Cloud power nodes cease strong tracking, but the energy storage state-of-charge corridor and operating boundary are retained, with short-term edge strategies dominating the fitting process; when insufficient energy storage reserves are anticipated, diesel generator preheating or parallel operation is triggered in advance. Long-term strategy deviation states continuously accumulate the deviation between the actual received values ​​of the equipment and the aligned cloud strategy, which can be used for limited compensation after communication is restored.

[0216] Example 4: Uninterrupted takeover of the network.

[0217] After communication is interrupted and then restored, the new cloud-based long-term strategy differs significantly from the current edge state. The edge sub-brain resets the cloud prediction starting point based on the actual measured state and limits the increment of cloud strategy weights in each control cycle; continuous power variables are progressively fused according to the fusion coefficient, and the diesel generator set start-up and shutdown states are switched only after the minimum start-up time, minimum shutdown time, and safety interlock conditions are met. Compared with direct switching, this avoids maximum power step jumps and equipment impact.

[0218] Example 5: Comparative simulation verification based on Simscape.

[0219] The electrical model in this embodiment directly adopts the RemoteMicrogrid model and its original equipment parameters from the official MathWorks Simscape example project, including the diesel generator set side, two sets of Battery Energy Storage Systems (BESS) and the average model converter, photovoltaic power station, distribution lines, transformer, main load, and F2 branch load (F2 is the original number of this branch load in the RemoteMicrogrid model). This invention only modifies the upper reference channel of the energy storage active power, without reconstructing the photovoltaic, energy storage, diesel generator set, or power electronic equipment models.

[0220] The cloud-based long-scale strategy update cycle is 1 second (corresponding to an actual timescale compression ratio of 1:3600), while the edge short-scale strategy and online control cycle is 5 ms. The SOC is generated in real time by an energy metering observer within the model based on the actual BESS active power, 150 kWh rated energy, and 80% of the initial values.

[0221] The simulation duration is 8 seconds with a fixed step size of 50 μs. Following the official case, the simulation includes a feeder switch action at 3 seconds and an F2 load shedding event at 6 seconds, with a communication interruption injected between 5.7 and 6.7 seconds. During the communication interruption, the cloud-based long-term strategy retains its last valid value; the edge-based auxiliary brain, after the communication interruption duration reaches 0.10 seconds, generates short-cycle corrections based solely on the island frequency and active power of the diesel generator set output in real time, while simultaneously updating the cloud-based and edge-based strategy weights, long-term strategy bias status, and compensation components.

[0222] The control group and the invention group used the exact same electrical model, initial state, event timing and solver settings. The only difference was the energy storage reference channel: the control group used the official original control channel, while the invention group used a control channel formed by cloud-based long-term time-scale strategy, edge short-term time-scale strategy, online fitting and cross-cycle compensation.

[0223] like Figures 10 to 22 The main indicators are statistically analyzed within the 5.8s to 7.5s event response window; SOC, power balance and graphics retain the complete simulation range.

[0224] Referring to Table 1, the key performance indicators of the control group and the present invention group are compared as follows:

[0225] Table 1

[0226] This invention controls the peak frequency deviation to decrease from 0.962405Hz to 0.915329Hz, a reduction of 4.89%; the average absolute frequency deviation decreases from 0.080174Hz to 0.066671Hz, a reduction of 16.84%. With this frequency improvement, the peak voltage deviation increases from 0.051956pu to 0.055570pu, and the BESS peak current increases from 0.401348pu to 0.424233pu, but the current within the event window remains below 1p.u.; the final SOC is approximately 79.91%, and the power balance residual RMS are 0.88687kW and 0.88659kW, respectively.

[0227] The results indicate that edge fast compensation achieves smaller frequency deviations during communication interruptions and load surges at the cost of moderately increasing transient voltage deviations and energy storage current utilization.

[0228] The dynamic process of cloud-edge dual-strategy fitting and bias adjustment is as follows: Figure 10 As shown, where Figure 10 Figure a shows the curves of the changes in the cloud-based long-scale strategy of the battery energy storage unit, the hold-up value during communication interruption, and the final fitted control quantity under the full time domain. Figure 10 Figure b shows the edge short-timescale correction curve at the corresponding moment. After the communication interruption is triggered at approximately 6 seconds, the cloud policy reference value enters a hold state. The fitted control quantity does not undergo a hard switch between the cloud policy and the edge policy, but rather smoothly transitions within the real-time feasible region as the weights are dynamically adjusted. The edge short-timescale correction quantity responds quickly to system power fluctuations, making real-time fine adjustments to the fitted control quantity. It can be seen that in communication anomaly scenarios, this method avoids the power impact caused by policy switching through continuous weight adjustment and constrained fitting, while ensuring the edge side's rapid response capability to short-term disturbances.

[0229] The interconnected changes of various state variables during communication interruption and recovery are as follows: Figure 12 As shown, where Figure 12 Figure a shows the effective communication indicator, which visually illustrates the complete process of the communication link switching from effective to ineffective in the interval of approximately 5.5s to 6.5s, and then regaining effectiveness. Figure 12 Figure b shows the duration of the communication interruption, which gradually accumulates over time after the communication failure and drops back to zero after communication is restored.

[0230] Figure 12 Figure c shows the dynamic change curves of cloud policy weight and edge policy weight. When communication is normal, the cloud policy weight remains at a high level, while the edge policy weight accounts for a low proportion. When communication is interrupted, the two types of weights do not undergo a step hard switch, but rather transition smoothly and continuously as communication reliability decreases. The cloud policy weight gradually decreases, while the edge policy weight gradually increases, ensuring the continuity of control commands.

[0231] Figure 12 The figure in the middle d-th diagram shows the change curves of the long-term strategy deviation state and the compensation component. During the communication interruption, the edge-side dominant control causes the actual running trajectory to deviate from the long-term strategy in the cloud, and the long-term strategy deviation state continues to accumulate. When communication is restored and the compensation triggering condition is met, the compensation component is gradually output, driving the long-term strategy deviation state to converge to zero, gradually reducing the strategy deviation accumulated during the interruption, and returning to the long-term optimization goal in the cloud.

[0232] It is evident that this method can achieve a smooth weight transition when communication deteriorates, while providing limited compensation for the policy deviation accumulated during the interruption, thus balancing short-term control stability with the consistency of long-term optimization objectives.

[0233] Load characteristics and source-side power allocation results under load abrupt change scenarios are as follows: Figure 13 As shown, where Figure 13 Figure a shows the curves of the main load, the load of branch F2, and the total load for each sample. The total load is composed of the superposition of the main load and the load of branch F2. At about 6 seconds, the load of branch F2 is cut off, and the total load drops in a step, which constitutes a typical load change condition. This is used to verify the dynamic response capability of the control method to source load disturbances.

[0234] Figure 13 Figure b shows the active power resource allocation curve under the control conditions of this invention. During the steady-state operation phase, the diesel generator set undertakes the basic load of the system, the battery energy storage unit smooths out small power fluctuations, and the photovoltaic power generation unit maintains a stable output. The three work together to match the total load demand. When a sudden load change occurs at 6 seconds, the battery energy storage unit adjusts its charging and discharging power first due to its rapid response characteristics, smoothing out the power imbalance caused by the load step. The diesel generator set smoothly adjusts its output within physical constraints such as ramp rate and power boundary, and the photovoltaic power generation unit maintains a stable output. Throughout the entire process, the power of each device does not jump, and the transition is always smooth within the real-time feasible domain.

[0235] It is evident that under load change conditions, this method can achieve smooth and coordinated power allocation from multiple sources through rapid correction and constrained fitting optimization of edge short-timescale strategies, which not only ensures system power balance but also avoids the impact of sudden changes in equipment commands.

[0236] Full-time variation of energy storage state of charge and comparison of schemes Figure 14 As shown, where Figure 14 Figure a shows the in-model observation curves of the state of charge of energy storage under the two control methods. During the entire time domain, the energy storage continues to discharge, and the state of charge shows a gradual downward trend. The overall change pattern of the two is consistent, indicating that the method does not deviate from the long-term energy operation trajectory of the system. Figure 14Figure b shows the difference in state of charge of energy storage between the control method of this invention and the original control method. When the 6-second load shedding event occurs, the difference fluctuates slightly and then gradually converges. The difference remains within a very small range throughout the process, which verifies that this method can ensure that the long-term energy state of energy storage remains consistent with the optimization target while responding quickly to disturbances.

[0237] Comparison of the dynamic response of system frequency, bus voltage and energy storage current, for example Figure 15 As shown, where Figure 15 Figure a shows the full-time frequency response from 0 to 8 seconds, including the system startup transient process; Figure 15 Figure b shows a local magnified view of the frequency response during communication interruption and load shedding. It can be seen that the frequency fluctuation amplitude under the control of the present invention is smaller, the recovery speed after disturbance is faster, and the frequency stability is better. Figure 15 Figure c shows the complete time-domain bus voltage response. During the transient process, the voltage deviation of this scheme increases slightly, but always remains within the acceptable range. Figure 15 The figure in the middle is the instantaneous maximum current curve of the three phases of the battery energy storage unit. This scheme improves the energy storage current utilization rate by smoothing out power fluctuations, but the current is lower than the 1p.u. limit throughout the process, which meets the equipment safety constraints.

[0238] The quantitative comparison results of key performance indicators are as follows: Figure 16 As shown, all indicators are relative to the original official control. The ratios of peak frequency deviation and average absolute frequency deviation are both less than 1. Under the control of this invention, the average absolute frequency deviation is significantly reduced, and the system frequency quality is clearly improved. The ratio of peak voltage deviation to energy storage peak current is slightly greater than 1, which is a reasonable trade-off for frequency stability, and the increase is within the acceptable range of the equipment and does not exceed the safety boundary.

[0239] The power tracking characteristics and compensation process of diesel generator sets are as follows: Figure 17 As shown, where Figure 17 Figure a shows the curves of the actual active power and reference power of the diesel generator set in the full time domain. The two curves have a high degree of matching and the power adjustment process is smooth without any step jumps. Figure 17 Figure b shows a magnified view of power tracking during communication and load events. After a sudden load change, the diesel generator set smoothly adjusts its output within the ramping constraint, without frequent start-stop actions. Figure 17 The middle diagram shows the dynamic changes in cloud policy weights and edge policy weights. After communication interruption, the weights transition continuously and smoothly without hard switching. Figure 17 The middle d-axis diagram shows the recovery component after communication is restored. It is output after the recovery conditions are met and gradually decays, driving the system to return to the long-term strategy trajectory in the cloud.

[0240] The multidimensional transient and steady-state characteristics of energy storage current, such as Figure 18 As shown, where Figure 18Figure a shows the instantaneous current envelope in the complete time domain from 0 to 8 seconds, which retains the starting peak value and shows that the current is stable after steady-state operation. Figure 18 Chinese B-map Figure 18 Figure C shows the effective and peak current values ​​of a non-overlapping 60Hz frequency. The current fluctuates slightly during the disturbance but remains below the 1p.u. limit throughout. Figure 18 d, Figure 18 e, Figure 18 f represents the original sampled waveforms of phases A, B, and C during the load shedding transient process. The three-phase waveforms are symmetrical, the transient transition is smooth, and there are no distortions or impact spikes, which verifies the power quality and equipment safety of the energy storage current under the control of this method.

[0241] The linkage process of multiple state variables within the communication interruption event window is as follows: Figure 19 As shown, the focus is on the core event range of 5.5 to 7 seconds. Figure 19 Figure a in the middle Figure 19 Figure b shows the effective communication flag and the duration of the communication interruption, visually presenting the event trigger moment; Figure 19 Figure c shows the cloud-based policy hold value and the final fitted control value. The fitted control value transitions smoothly throughout the process, without any power jumps caused by policy switching. Figure 19 The middle d-plot represents the edge short-timescale correction, which can quickly respond to local power disturbances; Figure 19 The figure in Figure e shows the smooth transition process of the weights of the two types of policies; Figure 19 The middle f figure shows the accumulation and recovery component output process of the long-term strategy deviation state. Figure 19 Chinese g-graph Figure 19 The middle and lower figures show a comparison of the physical frequency deviation and voltage deviation schemes. Under the control of this invention, the peak frequency deviation is smaller and the recovery is faster, which verifies the effectiveness of the entire logical closed loop from state perception to weight adjustment, control output, and deviation compensation.

[0242] The combined response of multiple electrical quantities in a battery energy storage unit, such as Figure 20 As shown, where Figure 20 Figure a in the middle Figure 20 Figure b shows the active and reactive power responses of the energy storage system. The active power smoothly adjusts with load fluctuations, while the reactive power quickly supports the bus voltage during transient processes. The two work together in coordination. Figure 20 Figure c shows the effective value of the non-overlapping 60Hz cycle current. The steady-state operation is stable, and the fluctuations are controllable during disturbances. Figure 20 The middle d-figure shows the observed state of charge of the energy storage. It shows a gradual downward trend throughout the time domain, which is consistent with the energy change law of the system. This indicates that the method can achieve rapid dynamic control while always adhering to the long-term energy constraints of energy storage.

[0243] The overall system's multi-source power distribution and power balancing effect is as follows: Figure 21 As shown, where Figure 21 Figure a shows the active power distribution curves of the total load, diesel generator set, and photovoltaic power generation unit, with the three types of power sources working together to match the load demand. Figure 21 Figure b shows a comparison of the active power schemes of the battery energy storage unit. Under the control of this invention, the energy storage regulation process is smoother and the impact is smaller. Figure 21 Figure c shows the reactive power resource allocation under the operating conditions of this invention. The diesel generator set, energy storage, and photovoltaics jointly provide reactive power support. During transient processes, energy storage undertakes the main reactive power regulation task due to its rapid response characteristics. Figure 21 The figure in the middle represents the system power balance residual. The residual level of this scheme is comparable to that of the original official control, and the steady-state accuracy is high, which verifies the power balance reliability and power supply quality under multi-source collaborative control.

[0244] Comparison of key performance indicators under different communication and load phases is as follows: Figure 22 As shown in the figure, the relative change rate of various indicators of the control method of this invention compared with the original official control method is presented in the form of a matrix heatmap. Negative values ​​indicate that the indicators are better than the original control method, and positive values ​​indicate that the indicators are higher than the original control method. The horizontal axis represents four typical operating conditions: before communication interruption, communication interruption and load not disconnected, communication interruption and load disconnected, and communication restored. The vertical axis represents five core performance indicators: mean absolute frequency difference, peak frequency difference, mean absolute voltage deviation, peak current of energy storage unit, and power balance residual error RMS.

[0245] Judging from the trend of indicator changes, the frequency quality indicators show an optimization trend throughout the process: the average absolute frequency difference is negative in all operating conditions, with the optimization reaching 27.1% in the communication recovery phase, which is the most significant improvement; the peak frequency difference decreases to varying degrees during the communication interruption, and only shows a slight increase in the communication recovery phase, with overall optimization still being the main trend.

[0246] There is a reasonable trade-off between voltage and current indicators: the average absolute voltage deviation increases significantly during the communication interruption and load disconnection phase, while it remains close to the original control level or slightly optimized during other phases; the peak current of the energy storage unit only increases slightly throughout the entire process, with a maximum increase of 6.7%, and remains within the safe operating range of the equipment.

[0247] The power balance accuracy remained basically consistent: the rate of change of the power balance residual RMS in all four stages was close to 0, and the maximum fluctuation did not exceed 3%, indicating that the control method of the present invention did not sacrifice the steady-state power balance accuracy of the system while optimizing dynamic performance.

[0248] Overall, this method significantly improves the frequency stability of the system at the cost of a small, acceptable increase in voltage and current. It can maintain excellent control performance, especially under harsh conditions such as communication interruption and sudden load changes, thus verifying the effectiveness and engineering applicability of the cloud-edge dual-strategy fitting and deviation compensation mechanism.

[0249] Example 6: Complete execution process under communication degradation.

[0250] At 10:00, the cloud-based mainframe generates a new version of the strategy package, containing reference values ​​and upper and lower boundaries for photovoltaic, energy storage, and diesel generator sets from 10:00 to 10:00 the next day. At 10:22, the edge-based auxiliary brain detects a continuous increase in packet loss rate and jitter, indicating that communication reliability is below the degradation threshold, and switches from state C0 to C1. When a node is lost at 10:30, the edge-based auxiliary brain reconstructs the continuous trajectory using the nodes at 10:15 and 10:45, the actual received values ​​of the equipment in the previous control cycle, and the ramp boundary. The available photovoltaic power on-site is lower than the cloud prediction, so the fitting optimization reduces the weight of the cloud strategy, increases the energy storage discharge command, and retains the diesel generator set for backup. The command is issued after being limited and protocol converted by the unified execution controller; the PCS returns that the available discharge power is derated to 80kW due to temperature, and the edge-based auxiliary brain shrinks the feasible domain for the next cycle accordingly. If the reliability recovers and continues to reach the exit threshold, the system returns to C0; if the communication interruption duration exceeds the communication interruption judgment time, it enters C2.

[0251] Example 7: Sudden drop in photovoltaic power and load step after communication interruption.

[0252] During communication interruptions, the edge sub-brain retains the energy storage state-of-charge corridor and the minimum start-stop boundary of the diesel generator set from the most recent policy package, but does not continue to strongly track expired power points. When local short-term source-load forecasts indicate insufficient future reserves, a diesel generator set preheating request is issued first; after a photovoltaic power outage, energy storage supports power balance on a rapid timescale, and the diesel generator set assumes continuous power after meeting the parallel operation conditions. The fitting optimization also limits the SOC from approaching the lower limit too quickly due to short-term support.

[0253] Example 8: The new strategy after network restoration conflicts with the on-site conditions.

[0254] When the network is restored, the new cloud-based long-term strategy assumes that the diesel generator sets are shut down, while the on-site diesel generator sets have already started operating due to increased load during the communication interruption. The edge computing system retains the on-site operating status, trims the new strategy to the current moment, and generates a progressively fused trajectory for continuous power variables. The diesel generator set shutdown is only executed after the minimum start-up time is met, energy storage reserves are sufficient, the predicted net load is lower than the diesel generator set exit threshold, and this condition persists for a preset time. The fused reference trajectory is then verified again by power balance, SOC, and ramp constraints.

[0255] Example 9: Predictive backfilling positioning.

[0256] When the preset replenishment conditions are not met and the long-term strategy deviation is in the accumulation phase, the edge auxiliary brain identifies future low-cost replenishment windows based on local short-term source load predictions, such as periods of surplus photovoltaic power, periods of low load, or economical operating periods of diesel generator sets. In the multiple control cycles preceding the arrival of the low-cost replenishment window, the edge auxiliary brain progressively pre-biases the state-of-charge (SOC) operating point of the energy storage units within the real-time feasible domain in a direction that reduces the long-term strategy deviation, thus distributing the replenishment component before the window. When the replenishment conditions are met, only the remaining small deviation needs to be executed, thereby reducing the impact of replenishment actions on the system's safety margin.

[0257] Example 10: Algorithm Alternative Method.

[0258] Long-term cloud-based strategies and short-term edge-based strategies can employ mathematical programming, model predictive control (MPC), rule-based optimization, or trained control model generation, respectively. Time alignment can utilize zero-order preservation, piecewise linearity, spline interpolation, or Hermite interpolation. Fitting optimization problems can be solved using quadratic programming, mixed-integer programming, explicit MPC, or hierarchical solutions. Alternative algorithms must maintain independent policies at both ends, joint weighting, constrained fitting, long-term policy bias state updates driven by actual device acceptance values, and constrained backfilling relationships.

[0259] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A cloud-edge dual-strategy fitting and deviation compensation control method for an off-grid photovoltaic-storage-diesel system, wherein the off-grid photovoltaic-storage-diesel system includes a cloud terminal, an edge terminal, a photovoltaic power generation unit, an energy storage unit, and a diesel generator set, characterized in that, Includes the following steps: S1: The cloud generates a long-term cloud strategy based on the first prediction time domain and sends the long-term cloud strategy to the edge terminal; the long-term cloud strategy includes reference power sequences of photovoltaic power generation units, energy storage units and diesel generator sets at multiple future time nodes; the edge terminal independently generates edge short-term strategies for photovoltaic power generation units, energy storage units and diesel generator sets based on the second prediction time domain and the actual measured state on site, where the first prediction time domain is longer than the second prediction time domain; S2: The edge maps the long-term cloud strategy to the timeline of the short-term edge strategy to obtain the cloud-aligned strategy. S3: Calculate communication reliability and net load mutation rate at the edge, and determine cloud policy weight and edge policy weight based on communication reliability and net load mutation rate; wherein, cloud policy weight increases with the increase of communication reliability, and edge policy weight increases with the increase of net load mutation rate; S4: At the edge, the alignment strategy with the cloud and the edge short-timescale strategy are used as parallel tracking targets. A fitting optimization problem is solved within the real-time feasible region to obtain a fitted control sequence. The objective function of the fitting optimization problem includes a first deviation term weighted by the cloud strategy weights and a second deviation term weighted by the edge strategy weights. The first deviation term is the deviation between the fitted control sequence and the alignment strategy with the cloud, and the second deviation term is the deviation between the fitted control sequence and the edge short-timescale strategy. The real-time feasible region includes power balance constraints, energy storage state of charge constraints, equipment power boundary constraints, ramping constraints, and diesel generator start-stop constraints. S5: The edge terminal sends the first control quantity in the fitted control sequence to the photovoltaic power generation unit, energy storage unit, and diesel generator set for execution, and collects the actual received value of the equipment. The actual received value of the equipment is the power value actually executed after the first control quantity is confirmed by the equipment controllers of the photovoltaic power generation unit, energy storage unit, and diesel generator set. The edge terminal uses the actual received value of the equipment as the starting control quantity for solving the fitting optimization problem in the next control cycle. S6: The edge device recursively updates the long-term policy deviation state based on the difference between the actual received value of the device and the reference power value of the cloud policy at the same time. When the communication reliability is higher than the compensation reliability threshold, the net load mutation degree is lower than the compensation mutation threshold, and the device safety margin meets the preset compensation conditions, the edge device generates a compensation component based on the long-term policy deviation state and incorporates the compensation component into the solution of the fitting optimization problem, gradually reducing the long-term policy deviation state under the constraints of the real-time feasible region.

2. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, The cloud-based long-term strategy also includes energy storage state of charge corridor, backup boundary, and device allowable boundary; when communication between the cloud and the edge is interrupted, the edge retains the energy storage state of charge corridor, backup boundary, and device allowable boundary as constraints of the real-time feasible domain.

3. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, The communication reliability is determined based on communication quality, policy timeliness, and state consistency. The communication quality is determined based on the average latency, packet loss rate, latency jitter, and the number of consecutive times a valid policy packet is not received within the sliding time window; The timeliness of the strategy is determined based on the time difference between the current moment and the generation time of the long-term strategy in the cloud. The state consistency is determined based on the deviation between the predicted state generated by the long-term cloud-based strategy and the actual measured state on-site.

4. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, The cloud policy weights and edge policy weights are determined through normalization calculations with positive zero-prevention parameters: ， ， in, The cloud strategy weight is the weight for the k-th control cycle; The edge policy weight for the kth control cycle; This is the baseline value for cloud-based policies; the baseline value for cloud-based policies increases with increasing communication reliability and decreases with increasing net load abruptness. This is the baseline value for the edge strategy; the baseline value for the edge strategy increases with the increase of the net load abrupt change. The parameter is a positive number to prevent zeroing.

5. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, The objective function of the fitting optimization problem also includes a control change penalty term, which is used to penalize the control change between adjacent control periods in the fitted control sequence.

6. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, The recursive update of the long-term strategy deviation state includes: multiplying the long-term strategy deviation state of the previous control cycle by a forgetting factor, and then superimposing the deviation increment obtained by time integration of the difference between the actual received value of the device and the reference power value of the cloud strategy at the same time, and limiting the superposition result within the preset allowable long-term strategy deviation state boundary.

7. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, The objective function of the fitting optimization problem also includes a penalty term for long-term policy deviation states.

8. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, After communication is restored, the edge device performs version verification and validity period verification on the new cloud long-term scale policy received, replaces the prediction starting point of the new cloud long-term scale policy with the actual measured status, and limits the increment of the cloud policy weight in each control cycle.

9. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 8, characterized in that, At the edge, continuous power variables in the reference power sequence of the new cloud long-term scale strategy are progressively fused according to a fusion coefficient, which is increased periodically by a preset increment until it equals 1. The start-stop state variables of the diesel generator set do not participate in the gradual fusion, and are switched only when the minimum start-up time, minimum shutdown time, and safety interlock conditions are met.

10. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, The replenishment component is generated according to the following formula: ， ， in, The compensation component is the component used in the k-th control cycle. This represents the long-term policy deviation state during the k-th control cycle. To compensate for the gain, To compensate for the gating coefficient, The communication reliability for the k-th control cycle, The net load mutation rate is given in the k-th control period. This refers to the equipment safety margin index for the k-th control cycle. The saturation function is defined to limit the backfill component within a preset change boundary, where ΔU is the preset change boundary.

11. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, When the compensation component is included in the solution of the fitting optimization problem, the net load mutation component in the compensation component is preferentially allocated to the energy storage unit and the photovoltaic power generation unit. The cross-cycle energy deviation component in the compensation component is compensated in multiple control cycles within the range allowed by the energy storage state of charge constraint and through the feasible start-stop window of the diesel generator set.

12. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, In S3, the edge also adjusts the cloud policy weight based on the modulus of the long-term policy deviation state. The cloud policy weight increases as the modulus of the long-term policy deviation state increases. When the net load mutation degree is higher than the compensation mutation threshold, the adjustment of the cloud policy weight based on the modulus of the long-term policy deviation state is suspended, and the cloud policy weight is determined based on the communication reliability and the net load mutation degree.

13. The cloud-edge dual-strategy fitting and deviation compensation control method for off-grid photovoltaic-storage-diesel system as described in claim 1, characterized in that, When the preset replenishment conditions are not met and the long-term strategy deviation is in the accumulation period, the edge end identifies the future low-cost replenishment window based on the local short-term source load prediction, and in the multiple control cycles before the arrival of the low-cost replenishment window, it gradually pre-biases the state of charge operation point of the energy storage unit in the direction that reduces the long-term strategy deviation within the real-time feasible domain.

14. The cloud-edge dual-strategy fitting and deviation compensation control method for an off-grid photovoltaic-storage-diesel system as described in claim 13, characterized in that, The low-cost replenishment window refers to the period of surplus photovoltaic power, the period of low load, or the economical operating period of diesel generator sets.

Images