Microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception

Abstract

The application discloses a micro-grid scheduling method based on causal probability wind power prediction and uncertainty perception, and comprises the following steps: obtaining historical wind power data, meteorological feature data, load data and unit operation constraint data; constructing a wind power probability prediction model and an input channel to generate quantile prediction values of wind power; generating predicted net load and predicted uncertainty features according to the quantile prediction values of wind power; constructing a reinforcement learning scheduling state, writing the predicted net load into the reinforcement learning scheduling state, and taking the current wind power and the predicted uncertainty features as auxiliary states; inputting the reinforcement learning scheduling state into a continuous action reinforcement learning agent to execute a scheduling strategy, and updating the continuous action reinforcement learning agent and the micro-grid operation state; and during the scheduling process, feeding back online update channel causal weights according to an operation evaluation quantity. The application can improve the utilization rate of prediction information and obtain a comprehensive optimal scheduling effect between wind power utilization rate and imbalance cost.

Description

Technical Field

[0001] This invention relates to the field of intelligent dispatching technology for new energy microgrids, specifically to a microgrid dispatching method based on causal probability wind power prediction and uncertainty perception. Background Technology

[0002] Wind power generation is characterized by randomness, intermittency, and volatility. For microgrids containing wind power, gas turbines, coal-fired power units, and adjustable loads, if the dispatch strategy only passively responds based on the current wind power generation, problems such as under-generation, over-generation, and increased wind curtailment may easily occur in the next moment.

[0003] Existing microgrid dispatching methods typically design prediction and dispatching models separately. One type of method uses point prediction of wind power and inputs the prediction results into the optimized dispatching model; another type of method uses reinforcement learning to directly learn the unit output and demand response actions. However, in short-term wind power dispatching, the current wind power has a strong autocorrelation with the wind power at the next time step. Simply attaching the predicted value to the end of the reinforcement learning state may cause the dispatching agent to learn a strategy of ignoring the predicted value, resulting in the small accuracy advantage of the prediction model not being effectively converted into dispatching benefits.

[0004] Other methods attempt to output wind power probability prediction intervals or confidence intervals, but these typically only use uncertainty as an evaluation metric for the prediction model, without establishing a state construction mechanism for scheduling decisions. If the scheduling state does not clearly distinguish between the predicted median and the prediction interval width, reinforcement learning agents struggle to simultaneously utilize wind power trends and risk levels.

[0005] Therefore, a method is needed that tightly couples causal probabilistic wind power forecasting with microgrid reinforcement learning scheduling, so that the predicted median enters the main signal slot of the scheduling state, the prediction interval width serves as an uncertainty auxiliary state, and reasonable scheduling is carried out in combination with microgrid physical constraints, wind curtailment and imbalance penalties. Summary of the Invention

[0006] The purpose of this invention is to provide a microgrid scheduling method based on causal probabilistic wind power prediction and uncertainty perception. This method can solve the problems of existing technologies, such as the prediction information being ignored by the reinforcement learning strategy, the difficulty in incorporating probabilistic prediction uncertainty into scheduling decisions, and the difficulty in coordinating and optimizing wind power consumption, in scenarios where wind power short-term prediction is linked with microgrid reinforcement learning scheduling.

[0007] To achieve the above functions, this invention designs a microgrid dispatching method based on causal probabilistic wind power prediction and uncertainty perception, executing the following steps S1-S8 to complete microgrid dispatching:

[0008] Step S1: Obtain historical wind power data, meteorological characteristic data, load data and unit operation constraint data in the microgrid, and arrange them into sequences according to time order. Divide the sequences into prediction model training segment, scheduling strategy training segment and scheduling strategy evaluation segment.

[0009] Step S2: Construct input channels, which include historical wind power channels and at least one meteorological feature channel, and input historical wind power data and meteorological feature data accordingly; calculate the statistical causal strength of each input channel for future wind power, and fuse the statistical causal strength with the physical prior weight to obtain the initial channel causal weight of each input channel;

[0010] Step S3: Construct a wind power probability prediction model. The wind power probability prediction model includes a causal channel gating module, a state-space temporal coding module, and a monotonic quantile output module. During the training phase of the prediction model, the causal channel gating module adjusts the input sequences of each input channel according to the channel causal weights, the state-space temporal coding module encodes the weighted sequence, and the monotonic quantile output module simultaneously outputs multiple ordered quantile prediction values ​​of wind power.

[0011] Step S4: Based on the median of the quantile predictions of multiple ordered wind power and the load at the next time moment, generate the predicted net load for the next time moment based on the load data, and generate the prediction uncertainty feature based on the interval width between at least two quantile predictions.

[0012] Step S5: Construct a reinforcement learning scheduling state, write the predicted net load of the next moment into the preset master signal slot in the reinforcement learning scheduling state that is directly related to wind power or net load, and use the current wind power and prediction uncertainty characteristics as auxiliary states.

[0013] Step S6: During the scheduling strategy training phase, the reinforcement learning scheduling state is input into the continuous action reinforcement learning agent, which generates normalized continuous actions. Based on the unit operation constraint data, an action boundary contraction coefficient is generated according to the prediction uncertainty characteristics. The action boundary contraction coefficient is used to dynamically adjust the controllable unit output adjustment range and the upper limit of the load node demand response ratio. The normalized continuous actions are mapped to the scheduling strategy and executed. The continuous action reinforcement learning agent is updated according to the next reinforcement learning scheduling state and operation evaluation quantity after the scheduling strategy is executed. The scheduling feedback quantity of each input channel is generated according to the operation evaluation quantity. The channel causal weights and the output bias of the causal channel gating module are updated online.

[0014] Step S7: Execute the scheduling strategy in the microgrid environment and update the microgrid operating status;

[0015] Step S8: Use the scheduling strategy evaluation section to evaluate the scheduling strategy, obtain the operation evaluation quantity, and complete the microgrid scheduling.

[0016] As a preferred technical solution of the present invention: in step S1, the prediction model training segment, the scheduling strategy training segment, and the scheduling strategy evaluation segment do not overlap in time, and the three are divided according to 60%, 20%, and 20% of the sequence.

[0017] As a preferred technical solution of the present invention, the specific steps of step S2 are as follows:

[0018] Step S2.1: For each meteorological characteristic channel, let the input data of the i-th input channel be denoted as . The PCMCI causality test method is used, and the significant connection strength obtained by the PCMCI causality test method is used as the statistical correlation of the i-th input channel with future wind power under lag. If the PCMCI causality test method is not available, the statistical correlation of the i-th input channel with future wind power under lag is calculated according to the following formula:

[0019] ;

[0020] in, For the i-th input channel in hysteresis The following is a statistical correlation of future wind power output; Indicates correlation statistics, This indicates that the i-th input channel is lagging at time t. The following input data, This represents the future wind power output h time steps ahead from time t. To predict the step size;

[0021] Step S2.2: Calculate the statistical causality strength of the i-th input channel. And calculate the normalized causality strength. :

[0022] ;

[0023] ;

[0024] in, For each input channel, calculate the maximum causal strength. Let j be the statistical causality strength of the j-th input channel. To prevent constants with zero denominators, D is the maximum lag order; Indicates each lagging lower The maximum value;

[0025] Step S2.3: Normalize the causal strength With physical prior weights Fusion, to obtain fusion weights The channel causal weights are obtained by linearly scaling the fusion weights:

[0026] ;

[0027] ;

[0028] in, The fusion coefficient is... , Let be the physical prior weight of the i-th input channel; Let be the fusion weight for the i-th input channel; Let be the channel causality weight of the i-th input channel; To preset the minimum channel causality weight, and These are the minimum and maximum values ​​of the fusion weights for each input channel, respectively; the channel causal weight for the historical wind power channel is set to 1.

[0029] As a preferred embodiment of the present invention: In step S3, the process of weighting the input channel by the causal channel gating module includes: denoting the input tensor as... The sliding window length is set to L; where m represents the total number of input channels, including one historical wind power channel and m-1 meteorological characteristic channels. This represents the input tensor of the 1st to mth input channels; the channel description vector is obtained by globally averaging the input tensors across the sliding window dimension for each input channel. And a gate vector is generated through a two-layer fully connected gated network. :

[0030] ;

[0031] The weighted input tensor is obtained by the following formula:

[0032] ;

[0033] in, Let i be the description vector of the i-th input channel, and let i be the channel description vector. , This represents the description vector for the first to mth input channels; For the gated vector, , This represents the gating coefficients for the 1st to mth input channels. Let be the gating coefficient for the i-th input channel; , and For gating network parameters; This represents the output bias, where, For the i-th output bias, Represents the sigmoid activation function; Represents the ReLU activation function; Let be the input tensor of the i-th input channel. The input tensor of the i-th input channel is weighted and adjusted; the state-space temporal coding module performs temporal coding on the weighted and adjusted input tensor to obtain the coded vector. .

[0034] As a preferred technical solution of the present invention: In step S3, the monotonic quantile output module is composed of the same encoding vector. Simultaneously output a baseline value and K-1 original increments K is the preset number of quantile levels, and Perform a nonnegative transformation on each original increment:

[0035] ;

[0036] The predicted quantile values ​​of the K wind power outputs are obtained using the following expression:

[0037] ;

[0038] ;

[0039] in, This is the k-th original increment; This represents an exponential function with the natural constant e as its base. It is a non-negative transformation function. The non-negative increment of the k-th quantile relative to the previous quantile. To find the non-negative increment corresponding to the sum index r; Let k be the predicted quantile value of the k-th wind power output. This is the quantile prediction value for the first wind power output; due to The predicted quantile values ​​of K wind power outputs do not decrease monotonically at the quantile level.

[0040] As a preferred embodiment of the present invention, the specific steps of step S4 are as follows:

[0041] Step S4.1: At the next time t+1, predict the median value. Total load at the next time point Combine and calculate the normalized predicted net load for the next time step. :

[0042] ;

[0043] in, This represents the peak load.

[0044] Step S4.2: Select the 95th percentile predicted value and 5th percentile predicted value Calculate the characteristics of prediction uncertainty:

[0045] ;

[0046] in, Corresponding to 90% of the interval width, This represents the uncertainty characteristic of the prediction at the next time step t+1.

[0047] As a preferred technical solution of the present invention: In step S5, the reinforcement learning scheduling state at time t... This includes the current output of controllable units, the output of controllable units at the previous moment, the load of each load node, time information, preset master signal slots, and residual supply-demand imbalance. For baseline agents that do not use prediction, the current wind power is written into the preset master signal slot. For agents that use point prediction, the predicted net load for the next moment is written into the preset master signal slot, and the current wind power is used as an auxiliary state. For agents that use probabilistic prediction, the predicted net load for the next moment is written into the preset master signal slot, and the current wind power and prediction uncertainty characteristics are used as auxiliary states.

[0048] As a preferred embodiment of the present invention: In step S6, the continuous action reinforcement learning agent includes an agent using point prediction and an agent using probability prediction, both of which will set the reinforcement learning scheduling state at time t. Input Policy Network To obtain the normalized continuous motion at time t , Here are the policy network parameters; Each component of the dimension lies in [-1,1]; The first M dimensions are mapped to the output setpoints of M controllable generator units:

[0049] ;

[0050] Will The N-dimensional mapping is represented by the demand response ratio of N load nodes:

[0051] ;

[0052] in, This represents the output setting value of the i-th controllable unit; for The corresponding action component of the i-th controllable unit. for The action component corresponding to the j-th load node, This represents the upper limit of the rated demand response ratio of the j-th load node; M and N are the number of controllable units and the number of load nodes, respectively, and the (M+j)-th dimension represents the j-th dimension in the last N dimensions; and These are the minimum and maximum outputs of the i-th controllable unit, respectively; The demand response ratio of the j-th load node;

[0053] Let the predictive uncertainty characteristic used by the current scheduling strategy be . The action boundary contraction coefficient is calculated using the following formula:

[0054] ;

[0055] in, The coefficient of contraction of the action boundary at time t; Represents the clipping function; To preset the shrinkage gain, ; To preset the minimum action boundary contraction coefficient, ;

[0056] The lower limit of dynamic output of the i-th controllable unit at time t and dynamic output limit They are respectively:

[0057] ;

[0058] ; in, The maximum ramp rate for the i-th controllable unit; The actual output of the i-th controllable unit at time t-1;

[0059] The upper limit of the dynamic demand response ratio of the j-th load node at time t. for:

[0060] ;

[0061] in, This represents the upper limit of the rated demand response ratio for the j-th load node;

[0062] When motion boundary adjustment is enabled, the rated output upper and lower limits are replaced with dynamic output upper and lower limits, and the normalized continuous motion is mapped to the output setpoint and demand response ratio:

[0063] ;

[0064] ;

[0065] in, This represents the output setpoint of the i-th controllable unit at time t; This represents the demand response ratio of the j-th load node at time t.

[0066] As a preferred embodiment of the present invention: In step S7, the microgrid environment performs constraint updates according to the following expression:

[0067] ;

[0068] ;

[0069] The load regulation constraints are:

[0070] ;

[0071] The output of the gas-balanced generator unit is:

[0072] ;

[0073] ;

[0074] in, This represents the intermediate output of the i-th controllable unit at time t; The actual output of the i-th controllable unit at time t. The actual output of the i-th controllable unit at time t-1. This represents the output setpoint of the i-th controllable unit at time t. Let be the maximum ramp rate of the i-th controllable unit. and Let be the minimum and maximum output of the i-th controllable unit, respectively. This represents the upper limit of the rated demand response ratio for the j-th load node; and Let represent the loads before and after adjustment at the j-th load node at time t, respectively. The total load after adjustment at time t; The total controllable unit output at time t; The intermediate output of the gas balance unit at time t; Let t be the actual output of the gas-balanced generator unit. This represents the actual output of the gas-balanced turbine unit at time t-1. This represents the maximum ramp rate for the gas-balanced generator unit. and These are the minimum and maximum outputs of the gas-balanced generator unit, respectively. Let t be the wind power received at time t;

[0075] Residual supply and demand imbalance at time t for:

[0076] ;

[0077] when At that time, reduce the wind power capacity accepted by the following expression:

[0078] ;

[0079] ;

[0080] in, Let t be the wind curtailment power. This indicates that an assignment update is being performed.

[0081] As a preferred technical solution of the present invention: the evaluation quantity in steps S6 and S8 is calculated using a technical evaluation function.

[0082] ;

[0083] in, This represents the technical evaluation function at time t. , , and To evaluate the weights, and These represent the imbalance penalty coefficients for under-generated and over-generated states, respectively; a continuous-action reinforcement learning agent is used to reduce... The scheduling policy is updated to achieve the target, and the scheduling policy evaluation segment is based on... Measure and evaluate dispatch strategies based on wind power acceptance capacity, wind curtailment capacity, and residual supply and demand imbalance;

[0084] For the k-th online update cycle, which contains H time points, let's denote the periodic average technical evaluation value when the gating coefficient is not disturbed. And add a preset disturbance to the gating coefficient of the i-th input channel. Post-period average technical evaluation They are respectively:

[0085] ;

[0086] in, This represents the technical evaluation function value at time t when the gate coefficients are not disturbed. This indicates that a preset perturbation is applied to the gating coefficient of the i-th input channel at time t. The subsequent technical evaluation function value;

[0087] Calculate the improvement in evaluation for the i-th input channel based on the periodic average technical evaluation before and after the disturbance:

[0088] ;

[0089] in, Let represent the evaluation improvement amount of the i-th input channel in the k-th online update cycle; normalize the evaluation improvement amount of each input channel to obtain the scheduling feedback amount of the i-th input channel:

[0090] ;

[0091] in, This represents the scheduling feedback quantity of the i-th input channel in the k-th online update cycle; This represents the minimum improvement in the evaluation of the input channel during the k-th online update cycle. This represents the maximum value of the evaluation improvement of the input channel in the k-th online update cycle; To prevent constants with a denominator of zero;

[0092] The channel causality weights from the previous online update cycle are weighted and fused with the scheduling feedback, and then the amplitude is limited to obtain the updated channel causality weights:

[0093] ;

[0094] in, This represents the channel causality weight of the i-th input channel in the (k+1)-th online update cycle. This represents the channel causality weight of the i-th input channel in the k-th online update cycle;

[0095] The output bias of the causal channel gating module is recalculated using the updated channel causal weights:

[0096] ;

[0097] in, This represents the output bias of the i-th input channel of the causal channel gating module in the (k+1)-th online update cycle. This is the set of times corresponding to the kth online update cycle; This is the preset disturbance amount for the gating coefficient; To preset the online update step size, ; Preset minimum channel causality weight; To prevent constants with a denominator of zero.

[0098] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0099] 1. This invention converts the median wind power into predicted net load and writes it into a preset slot in the reinforcement learning state that originally carried the main information of wind power, so that the scheduling strategy can directly use the predicted information at the most sensitive wind power or net load state position, instead of just passively receiving the predicted value in the additional dimension.

[0100] 2. This invention uses the prediction interval width as input to the reinforcement learning agent to differentiate between low-risk and high-risk predictions, enabling the scheduling strategy to adjust the output and demand response of coal-fired units in high-uncertainty scenarios, thereby improving scheduling robustness. Furthermore, by utilizing the uncertainty of prediction to adjust the action boundaries, the range of unit output variation and the upper limit of load response ratio are narrowed in high-uncertainty scenarios, reducing the operational risks caused by aggressive adjustments.

[0101] 3. This invention integrates PCMCI statistical causal strength with physical prior weights, giving the channel gating of the wind power prediction model both data-driven and physical prior basis. Through a monotonic quantile output structure, multiple quantiles naturally satisfy a monotonic non-decreasing relationship, reducing the impact of quantile crossovers on the prediction interval and scheduling state construction. The scheduling evaluation results are further fed back to the causal channel gating module, enabling the channel causal weights to be corrected online according to operating conditions, forming a closed loop of mutual feedback between prediction and scheduling.

[0102] 4. This invention constructs a closed-loop training and evaluation environment that meets the microgrid dispatching objectives by setting the gas turbine unit as an automatic balancing unit and entrusting the coal-fired unit and demand response to reinforcement learning control.

[0103] 5. This invention forms a complete link from data isolation, causal probability prediction, state interface design, reinforcement learning scheduling, microgrid physical execution to multi-index evaluation, which can be used in day-ahead, intraday or short-term rolling scheduling scenarios of microgrids with a high proportion of wind power. Attached Figure Description

[0104] Figure 1 This is a structural diagram of a microgrid dispatching method based on causal probabilistic wind power prediction and uncertainty perception provided in an embodiment of the present invention;

[0105] Figure 2 This is a flowchart of a microgrid dispatching method based on causal probabilistic wind power prediction and uncertainty perception provided in an embodiment of the present invention;

[0106] Figure 3 This is a schematic diagram of a wind power probability prediction model provided according to an embodiment of the present invention;

[0107] Figure 4 This is a schematic diagram of the scheduling state construction provided in an embodiment of the present invention;

[0108] Figure 5 This is a schematic diagram of a microgrid reinforcement learning scheduling environment provided according to an embodiment of the present invention. Detailed Implementation

[0109] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0110] The microgrid dispatching method based on causal probabilistic wind power prediction and uncertainty perception provided in this embodiment of the invention refers to... Figure 1 , Figure 2 Perform the following steps S1-S8 to complete the microgrid dispatch:

[0111] Step S1: Obtain historical wind power data, meteorological characteristic data, load data and unit operation constraint data in the microgrid, and arrange them into sequences according to time order. Divide the sequences into prediction model training segment, scheduling strategy training segment and scheduling strategy evaluation segment.

[0112] In this embodiment, historical wind power data and meteorological characteristic data with a 15-minute resolution are collected. The meteorological characteristic data includes wind speed, wind direction, temperature, humidity, and air pressure. The prediction model training segment, scheduling strategy training segment, and scheduling strategy evaluation segment do not overlap in time. The three segments are divided into 60%, 20%, and 20% of the sequence, respectively, so that the wind power probability prediction model is trained only on the first 60% of the data.

[0113] Step S2: Construct input channels, which include historical wind power channels and at least one meteorological feature channel, and input historical wind power data and meteorological feature data accordingly; calculate the statistical causal strength of each input channel for future wind power, and fuse the statistical causal strength with the physical prior weight to obtain the initial channel causal weight of each input channel;

[0114] In this embodiment, a sliding window of length 96 is used to construct the prediction sample. The window includes historical wind power and five meteorological feature channels, corresponding to wind speed, wind direction, temperature, humidity, and air pressure, respectively. The length of 96 corresponds to a one-day time window at a 15-minute resolution. In the software implementation, the `build_windows` function can concatenate the historical wind power data sequence and the meteorological feature data sequence into a time-series input.

[0115] The specific steps of step S2 are as follows:

[0116] Step S2.1: For each meteorological characteristic channel, record the input data of the i-th input channel. The PCMCI causality test method is used, and the significant connection strength obtained by the PCMCI causality test method is used as the statistical correlation of the i-th input channel with future wind power under lag. If the PCMCI causality test method is not available, the statistical correlation of the i-th input channel with future wind power under lag is calculated according to the following formula:

[0117] ;

[0118] in, For the i-th input channel in hysteresis The following is a statistical correlation of future wind power output; Indicates correlation statistics, This indicates that the i-th input channel is lagging at time t. The following input data, This represents the future wind power output h time steps ahead from time t. The prediction step size is a positive integer. In the example, ;

[0119] Step S2.2: Calculate the statistical causality strength of the i-th input channel. And calculate the normalized causality strength. :

[0120] ;

[0121] ;

[0122] in, For each input channel, calculate the maximum causal strength. Let j be the statistical causality strength of the j-th input channel. To prevent constants with zero denominators, D is the maximum lag order; Indicates each lagging lower The maximum value;

[0123] Step S2.3: Normalize the causal strength With physical prior weights In the fusion process, physical prior weights assign higher prior weights to historical wind power and wind speed channels than to wind direction, temperature, humidity, and air pressure channels; thus, the fusion weights are obtained. The channel causal weights are obtained by linearly scaling the fusion weights:

[0124] ;

[0125] ;

[0126] in, The fusion coefficient is... , Let be the physical prior weight of the i-th input channel; Let be the fusion weight for the i-th input channel; Let be the channel causality weight of the i-th input channel; To preset the minimum channel causality weight, and These are the minimum and maximum values ​​of the fusion weights for each input channel, respectively; the channel causal weight for the historical wind power channel is set to 1.

[0127] In the software implementation of this embodiment, the channel causal weights are calculated using the `run_pcmci` function. The PCMCI method is invoked preferentially, calculating the significant causal strength of each input channel on future wind power within the lag range of `tau_min=1` to `PCMCI_TAU_MAX=4`, with a significance threshold of `PCMCI_ALPHA=0.05`. When PCMCI dependency is unavailable, the Pearson lag correlation between the meteorological characteristic channel and the wind power at the next time step is used as the degradation strength.

[0128] In the software implementation, the physical prior weights are set to wind speed 1.00, wind direction 0.50, temperature 0.30, humidity 0.20, and air pressure 0.35. The fusion weight between the statistical causal strength and the physical prior weights is FUSION_ALPHA=0.5. The fused weights are scaled according to the minimum channel weight CAUSAL_MIN_W=0.10 and concatenated with the historical wind power channel weight of 1.0 to obtain the final channel causal weights causal_weights.

[0129] Step S3: Construct a wind power probability prediction model. The wind power probability prediction model includes a causal channel gating module, a state-space temporal coding module, and a monotonic quantile output module. During the training phase of the prediction model, the causal channel gating module adjusts the input sequences of each input channel according to the channel causal weights, the state-space temporal coding module encodes the weighted sequence, and the monotonic quantile output module simultaneously outputs multiple ordered quantile prediction values ​​of wind power.

[0130] Reference Figure 3 In step S3, the process of the causal channel gating module weighting the input channel includes: denoting the input tensor as... The sliding window length is set to L; where m represents the total number of input channels, including one historical wind power channel and m-1 meteorological characteristic channels. This represents the input tensor of the 1st to mth input channels; the channel description vector is obtained by globally averaging the input tensors across the sliding window dimension for each input channel. And a gate vector is generated through a two-layer fully connected gated network. :

[0131] ;

[0132] The weighted input tensor is obtained by the following formula:

[0133] ;

[0134] in, Let i be the description vector of the i-th input channel, and let i be the channel description vector. , This represents the description vector for the first to mth input channels; For the gated vector, , This represents the gating coefficients for the 1st to mth input channels. Let be the gating coefficient for the i-th input channel; , and For gating network parameters; This represents the output bias, where, For the i-th output bias, Represents the sigmoid activation function; Represents the ReLU activation function; Let be the input tensor of the i-th input channel. The input tensor of the i-th input channel is weighted and adjusted; the state-space temporal coding module performs temporal coding on the weighted and adjusted input tensor to obtain the coded vector. .

[0135] In step S3, the monotonic quantile output module is composed of the same encoding vector. Simultaneously output a baseline value and K-1 original increments K is the preset number of quantile levels, and Perform a nonnegative transformation on each original increment:

[0136] ;

[0137] The predicted quantile values ​​of the K wind power outputs are obtained using the following expression:

[0138] ;

[0139] ;

[0140] in, This is the k-th original increment; This represents an exponential function with base e. It is a non-negative transformation function. The non-negative increment of the k-th quantile relative to the previous quantile. To find the non-negative increment corresponding to the sum index r; Let k be the predicted quantile value of the k-th wind power output. This is the quantile prediction value for the first wind power output; due to The predicted quantile values ​​of K wind power outputs do not decrease monotonically at the quantile level.

[0141] In this embodiment, the monotonic quantile output module outputs at least three quantile prediction values ​​of wind power from q05, q10, q25, q50, q75, q90, and q95.

[0142] In the software implementation of this embodiment, the wind power probabilistic prediction model can call CGProbMambaModel. This model first adjusts the input tensor to a channel-priority form and performs causal channel gating through CausalSEAttention; then, the channel-weighted sequence is input into a linear projection layer and a multi-layer Mamba state-space temporal coding layer; finally, the hidden state of the last time step is taken, normalized, and input into MonotonicQuantileHead to output multiple quantile prediction values ​​of wind power.

[0143] In the software implementation, CausalSEAttention registers the channel causal weights as a buffer tensor `causal_prior` and initializes the gate layer bias with these causal weights, ensuring that the channel gating has a priori physical causality from the early stages of training. CausalSEAttention performs compression, nonlinear transformation, and sigmoid gating on the input channels, outputting dynamic weights for each channel and scaling the input channels. During online operation, the causal channel gating module receives scheduling evaluation feedback according to a preset online update cycle. The process includes calculating the evaluation improvement amount for online channel causal weight updates, normalizing the scheduling feedback amount, weighted fusion, and recalculating the output bias.

[0144] In the software implementation, MonotonicQuantileHead outputs n_quantiles values, where the 0th value is the baseline and the remaining values ​​are the original increments. The original increments are then subjected to a softplus non-negative transformation and accumulated onto the baseline, forming a monotonically non-decreasing quantile sequence. The quantile levels can be [0.05, 0.10, 0.25, 0.50, 0.75, 0.90, 0.95], where the median MEDIAN_IDX corresponds to the 0.50 quantile.

[0145] In this embodiment, the wind power probability prediction model is trained using wind power quantile loss; for quantile levels Actual wind power output and predicted quantiles Wind power quantile loss Represented as:

[0146] ;

[0147] The average of multiple wind power quantile losses is calculated by batch and wind power quantile dimension.

[0148] Step S4: Based on the median of the quantile predictions of multiple ordered wind power and the load at the next time moment, generate the predicted net load for the next time moment based on the load data, and generate the prediction uncertainty feature based on the interval width between at least two quantile predictions.

[0149] Reference Figure 4 The specific steps of step S4 are as follows:

[0150] Step S4.1: At the next time t+1, predict the median value. Total load at the next time point Combine and calculate the normalized predicted net load for the next time step. :

[0151] ;

[0152] in, This represents the peak load.

[0153] Step S4.2: Select the 95th percentile predicted value and 5th percentile predicted value Calculate the characteristics of prediction uncertainty:

[0154] ;

[0155] in, Corresponding to 90% of the interval width, This represents the uncertainty characteristic of the prediction at the next time step t+1.

[0156] In the embodiment, the difference between the 90th percentile predicted value q90 and the 10th percentile predicted value q10 can also be selected as the prediction interval width.

[0157] In the software implementation of the example, the wind power probability prediction model outputs a data file for reinforcement learning. In the software implementation, the process_pred_prob function converts the data file into MW units and generates the file pred_cgprob_{segment_name}.csv.

[0158] Step S5: Construct a reinforcement learning scheduling state, write the predicted net load of the next moment into the preset master signal slot in the reinforcement learning scheduling state that is directly related to wind power or net load, and use the current wind power and prediction uncertainty characteristics as auxiliary states.

[0159] In step S5, the reinforcement learning scheduling state at time t This includes the current output of controllable units, the output of controllable units at the previous moment, the load of each load node, time information, preset master signal slots, and residual supply-demand imbalance. For baseline agents that do not use prediction, the current wind power is written into the preset master signal slot. For agents that use point prediction, the predicted net load for the next moment is written into the preset master signal slot, and the current wind power is used as an auxiliary state. For agents that use probabilistic prediction, the predicted net load for the next moment is written into the preset master signal slot, and the current wind power and prediction uncertainty characteristics are used as auxiliary states.

[0160] Step S6: During the scheduling strategy training phase, the reinforcement learning scheduling state is input into the continuous action reinforcement learning agent, which generates normalized continuous actions. Based on the unit operation constraint data, an action boundary contraction coefficient is generated according to the prediction uncertainty characteristics. The action boundary contraction coefficient is used to dynamically adjust the controllable unit output adjustment range and the upper limit of the load node demand response ratio. The normalized continuous actions are mapped to the scheduling strategy and executed. The continuous action reinforcement learning agent is updated according to the next reinforcement learning scheduling state and operation evaluation quantity after the scheduling strategy is executed. The scheduling feedback quantity of each input channel is generated according to the operation evaluation quantity. The channel causal weights and the output bias of the causal channel gating module are updated online.

[0161] The dispatch strategy includes the output setpoint of at least one controllable unit (coal-fired unit) and the demand response ratio of at least one load node. Gas-fired units can act as automatic balancing units, without directly entering the reinforcement learning action space, but are automatically compensated by the microgrid environment based on load, wind power acceptance, and the output deficit of coal-fired units.

[0162] Reference Figure 5 In step S6, the continuous action reinforcement learning agent includes an agent using point prediction and an agent using probability prediction. Both agents will use the reinforcement learning scheduling state at time t. Input Policy Network To obtain the normalized continuous motion at time t , Here are the policy network parameters; Each component of the dimension lies in [-1,1]; The first M dimensions are mapped to the output setpoints of M controllable generator units:

[0163] ;

[0164] Will The N-dimensional mapping is represented by the demand response ratio of N load nodes:

[0165] ;

[0166] in, This represents the output setting value of the i-th controllable unit; for The corresponding action component of the i-th controllable unit. for The action component corresponding to the j-th load node, This represents the upper limit of the rated demand response ratio of the j-th load node; M and N are the number of controllable units and the number of load nodes, respectively, and the (M+j)-th dimension represents the j-th dimension in the last N dimensions; and These are the minimum and maximum outputs of the i-th controllable unit, respectively; Let be the demand response ratio of the j-th load node.

[0167] Let the predictive uncertainty characteristic used by the current scheduling strategy be . The action boundary contraction coefficient is calculated using the following formula:

[0168] ;

[0169] in, The coefficient of contraction of the action boundary at time t; Represents the clipping function; To preset the shrinkage gain, ; To preset the minimum action boundary contraction coefficient, ;

[0170] The lower limit of dynamic output of the i-th controllable unit at time t and dynamic output limit They are respectively:

[0171] ;

[0172] ; in, The maximum ramp rate for the i-th controllable unit; The actual output of the i-th controllable unit at time t-1; The smaller the value, the smaller the dynamic output adjustment range and the upper limit of the dynamic demand response ratio;

[0173] The upper limit of the dynamic demand response ratio of the j-th load node at time t. for:

[0174] ;

[0175] in, This represents the upper limit of the rated demand response ratio for the j-th load node;

[0176] When motion boundary adjustment is enabled, the rated output upper and lower limits are replaced with dynamic output upper and lower limits, and the normalized continuous motion is mapped to the output setpoint and demand response ratio:

[0177] ;

[0178] ;

[0179] in, This represents the output setpoint of the i-th controllable unit at time t; This represents the demand response ratio of the j-th load node at time t.

[0180] Step S7: Execute the scheduling strategy in the microgrid environment and update the microgrid operating status;

[0181] In step S7, the microgrid environment performs constraint updates according to the following expression:

[0182] ;

[0183] ;

[0184] The load regulation constraints are:

[0185] ;

[0186] The output of the gas-balanced generator unit is:

[0187] ;

[0188] ;

[0189] in, This represents the intermediate output of the i-th controllable unit at time t; The actual output of the i-th controllable unit at time t. The actual output of the i-th controllable unit at time t-1. This represents the output setpoint of the i-th controllable unit at time t. Let be the maximum ramp rate of the i-th controllable unit. and Let be the minimum and maximum output of the i-th controllable unit, respectively. This represents the upper limit of the rated demand response ratio for the j-th load node; and Let represent the loads before and after adjustment at the j-th load node at time t, respectively. The total load after adjustment at time t; The total controllable unit output at time t; The intermediate output of the gas balance unit at time t; Let t be the actual output of the gas-balanced generator unit. This represents the actual output of the gas-balanced turbine unit at time t-1. This represents the maximum ramp rate for the gas-balanced generator unit. and These are the minimum and maximum outputs of the gas-balanced generator unit, respectively. Let t be the wind power received at time t;

[0190] Residual supply and demand imbalance at time t for:

[0191] ;

[0192] when At that time, reduce the wind power capacity accepted by the following expression:

[0193] ;

[0194] ;

[0195] in, Let t be the wind curtailment power. This indicates that an assignment update is being performed.

[0196] Step S8: Use the scheduling strategy evaluation section to evaluate the trained scheduling strategy, obtain the operation evaluation quantity, and complete the microgrid scheduling.

[0197] In steps S6 and S8, the evaluation metrics are calculated using a technical evaluation function.

[0198] ;

[0199] in, This represents the technical evaluation function at time t. , , and To evaluate the weights, and These represent the imbalance penalty coefficients for under-generated and over-generated states, respectively; a continuous-action reinforcement learning agent is used to reduce... The scheduling policy is updated to achieve the target, and the scheduling policy evaluation segment is based on... Measure and evaluate dispatch strategies based on wind power acceptance capacity, wind curtailment capacity, and residual supply and demand imbalance;

[0200] For the k-th online update cycle, which contains H time points, let's denote the periodic average technical evaluation value when the gating coefficient is not disturbed. And add a preset disturbance to the gating coefficient of the i-th input channel. Post-period average technical evaluation They are respectively:

[0201] ;

[0202] in, This represents the technical evaluation function value at time t when the gate coefficients are not disturbed. This indicates that a preset perturbation is applied to the gating coefficient of the i-th input channel at time t. The subsequent technical evaluation function value;

[0203] Calculate the improvement in evaluation for the i-th input channel based on the periodic average technical evaluation before and after the disturbance:

[0204] ;

[0205] in, Let represent the evaluation improvement amount of the i-th input channel in the k-th online update cycle; normalize the evaluation improvement amount of each input channel to obtain the scheduling feedback amount of the i-th input channel:

[0206] ;

[0207] in, This represents the scheduling feedback quantity of the i-th input channel in the k-th online update cycle; This represents the minimum improvement in the evaluation of the input channel during the k-th online update cycle. This represents the maximum value of the evaluation improvement of the input channel in the k-th online update cycle; To prevent constants with a denominator of zero;

[0208] The channel causality weights from the previous online update cycle are weighted and fused with the scheduling feedback, and then the amplitude is limited to obtain the updated channel causality weights:

[0209] ;

[0210] in, This represents the channel causality weight of the i-th input channel in the (k+1)-th online update cycle. This represents the channel causality weight of the i-th input channel in the k-th online update cycle;

[0211] The output bias of the causal channel gating module is recalculated using the updated channel causal weights:

[0212] ;

[0213] in, This represents the output bias of the i-th input channel of the causal channel gating module in the (k+1)-th online update cycle. This is the set of times corresponding to the kth online update cycle; This is the preset disturbance amount for the gating coefficient; To preset the online update step size, ; Preset minimum channel causality weight; To prevent constants with zero denominators; when the evaluation quantity decreases after perturbing the i-th input channel, It is positive.

[0214] In this embodiment, the microgrid core environment can be implemented using WindMicrogridCore. The reinforcement learning environment can be implemented using WindMicrogridEnv. The reinforcement learning action space is a 6-dimensional continuous action, including the output setpoints of three coal-fired units and the demand response ratios of three load nodes. The maximum output of the coal-fired units can be 50MW, 200MW, and 80MW, respectively, and the ramping constraints can be 12MW, 35MW, and 18MW per step, respectively. The demand response ratio can be mapped to the load transfer of each load node within a predetermined range. In the uncertainty-aware action boundary adjustment, the higher the prediction uncertainty, the smaller the action boundary contraction coefficient; when the prediction uncertainty is within a preset low value range, the action boundary contraction coefficient is 1, restoring the rated unit output range and the upper limit of the rated demand response ratio.

[0215] In the software implementation of this embodiment, the training parameters for the continuous action reinforcement learning agent may include a learning rate of 5e-5, an experience replay capacity of 200,000, a batch size of 1024, a discount factor gamma = 0.995, a soft update coefficient tau = 0.005, a learning start step of 5000, a policy network structure of [400, 300], and an entropy coefficient ent_coef = 'auto_0.05'. These parameters are not the sole limitation of the scope of this invention and can be adjusted according to the microgrid scale and data sampling interval.

[0216] In the software implementation, the RewardLogger records the average reward within a specified window during the training process. When the average reward exceeds the historical best value, model.zip is saved as the optimal model, and model_final.zip is saved at the end of training. This optimal model checkpoint mechanism is used to reduce the impact of performance degradation in the later stages of continuous action reinforcement learning training on scheduling evaluation.

[0217] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception, characterized in that, Perform the following steps S1-S8 to complete the microgrid dispatch: Step S1: Obtain historical wind power data, meteorological characteristic data, load data and unit operation constraint data in the microgrid, and arrange them into sequences according to time order. Divide the sequences into prediction model training segment, scheduling strategy training segment and scheduling strategy evaluation segment. Step S2: Construct input channels, which include historical wind power channels and at least one meteorological feature channel, and input historical wind power data and meteorological feature data accordingly; calculate the statistical causal strength of each input channel for future wind power, and fuse the statistical causal strength with the physical prior weight to obtain the initial channel causal weight of each input channel; Step S3: Construct a wind power probability prediction model. The wind power probability prediction model includes a causal channel gating module, a state-space temporal coding module, and a monotonic quantile output module. During the training phase of the prediction model, the causal channel gating module adjusts the input sequences of each input channel according to the channel causal weights, the state-space temporal coding module encodes the weighted sequence, and the monotonic quantile output module simultaneously outputs multiple ordered quantile prediction values ​​of wind power. Step S4: Based on the median of the quantile predictions of multiple ordered wind power and the load at the next time moment, generate the predicted net load for the next time moment based on the load data, and generate the prediction uncertainty feature based on the interval width between at least two quantile predictions. Step S5: Construct a reinforcement learning scheduling state, write the predicted net load of the next moment into the preset master signal slot in the reinforcement learning scheduling state that is directly related to wind power or net load, and use the current wind power and prediction uncertainty characteristics as auxiliary states. Step S6: During the scheduling strategy training phase, the reinforcement learning scheduling state is input into the continuous action reinforcement learning agent, which generates normalized continuous actions. Based on the unit operation constraint data, the action boundary contraction coefficient is generated according to the prediction uncertainty characteristics. The action boundary contraction coefficient is used to dynamically adjust the controllable unit output adjustment range and the upper limit of the load node demand response ratio. The normalized continuous actions are mapped to the scheduling strategy and executed. The continuous action reinforcement learning agent is updated based on the next reinforcement learning scheduling state and the running evaluation after the execution of the scheduling strategy, and the scheduling feedback of each input channel is generated based on the running evaluation. The channel causal weights and the output bias of the causal channel gating module are updated online. Step S7: Execute the scheduling strategy in the microgrid environment and update the microgrid operating status; Step S8: Use the scheduling strategy evaluation section to evaluate the scheduling strategy, obtain the operation evaluation quantity, and complete the microgrid scheduling.

2. The microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception according to claim 1, characterized in that, In step S1, the prediction model training segment, the scheduling strategy training segment, and the scheduling strategy evaluation segment do not overlap in time, and the three are divided according to 60%, 20%, and 20% of the sequence.

3. The microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step S2.1: For each meteorological characteristic channel, let the input data of the i-th input channel be denoted as . The PCMCI causality test method is used, and the significant connection strength obtained by the PCMCI causality test method is used as the statistical correlation of the i-th input channel with future wind power under lag. If the PCMCI causality test method is not available, the statistical correlation of the i-th input channel with future wind power under lag is calculated according to the following formula: ； in, For the i-th input channel in hysteresis The following is a statistical correlation of future wind power output; Indicates correlation statistics, This indicates that the i-th input channel is lagging at time t. The following input data, Indicates starting from time t and moving forward... The future wind power capacity at this time step To predict the step size; Step S2.2: Calculate the statistical causality strength of the i-th input channel. And calculate the normalized causality strength. : ； ； in, For each input channel, calculate the maximum causal strength. Let j be the statistical causality strength of the j-th input channel. To prevent constants with zero denominators, D is the maximum lag order; Indicates each lagging lower The maximum value; Step S2.3: Normalize the causal strength The fused weights are obtained by fusing them with the physical prior weights, and then the channel causality weights are obtained by linearly scaling the fused weights. ； ； in, The fusion coefficient is... , Let be the physical prior weight of the i-th input channel; Let be the fusion weight for the i-th input channel; Let be the channel causality weight of the i-th input channel; To preset the minimum channel causality weight, and These are the minimum and maximum values ​​of the fusion weights for each input channel, respectively; the channel causal weight for the historical wind power channel is set to 1.

4. The microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception according to claim 3, characterized in that, In step S3, the process of the causal channel gating module weighting the input channel includes: denoting the input tensor as... The sliding window length is set to L; where m represents the total number of input channels, including one historical wind power channel and m-1 meteorological characteristic channels. This represents the input tensor of the 1st to mth input channels; the channel description vector is obtained by globally averaging the input tensors across the sliding window dimension for each input channel. And a gate vector is generated through a two-layer fully connected gated network. : ； The weighted input tensor is obtained by the following formula: ； in, Let i be the description vector of the i-th input channel, and let i be the channel description vector. , This represents the description vector for the first to mth input channels; For the gated vector, , This represents the gating coefficients for the 1st to mth input channels. Let be the gating coefficient for the i-th input channel; , and For gating network parameters; This represents the output bias, where, For the i-th output bias, Represents the sigmoid activation function; Represents the ReLU activation function; Let be the input tensor of the i-th input channel. The input tensor of the i-th input channel is weighted and adjusted; the state-space temporal coding module performs temporal coding on the weighted and adjusted input tensor to obtain the coded vector. .

5. The microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception according to claim 4, characterized in that, In step S3, the monotonic quantile output module is composed of the same encoding vector. Simultaneously output a baseline value and K-1 original increments K is the preset number of quantile levels, and ; Perform a nonnegative transformation on each original increment: ； The predicted quantile values ​​of the K wind power outputs are obtained using the following expression: ； ； in, This is the k-th original increment; This represents an exponential function with the natural constant e as its base. It is a non-negative transformation function. The non-negative increment of the k-th quantile relative to the previous quantile. To find the non-negative increment corresponding to the sum index r; Let k be the predicted quantile value of the k-th wind power output. This is the quantile prediction value for the first wind power output; due to The predicted quantile values ​​of K wind power outputs do not decrease monotonically at the quantile level.

6. The microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception according to claim 5, characterized in that, The specific steps of step S4 are as follows: Step S4.1: At the next time t+1, predict the median value. Total load at the next time point Combine and calculate the normalized predicted net load for the next time step. : ； in, This represents the peak load. Step S4.2: Select the 95th percentile predicted value and 5th percentile predicted value Calculate the characteristics of prediction uncertainty: ； in, Corresponding to 90% of the interval width, This represents the uncertainty characteristic of the prediction at the next time step t+1.

7. The microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception according to claim 6, characterized in that, In step S5, the reinforcement learning scheduling state at time t This includes the current output of controllable units, the output of controllable units at the previous moment, the load of each load node, time information, preset master signal slots, and residual supply-demand imbalance. For baseline agents that do not use prediction, the current wind power is written into the preset master signal slot. For agents that use point prediction, the predicted net load for the next moment is written into the preset master signal slot, and the current wind power is used as an auxiliary state. For agents that use probabilistic prediction, the predicted net load for the next moment is written into the preset master signal slot, and the current wind power and prediction uncertainty characteristics are used as auxiliary states.

8. The microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception according to claim 7, characterized in that, In step S6, the continuous action reinforcement learning agent includes an agent using point prediction and an agent using probability prediction. Both agents will set the reinforcement learning scheduling state at time t. Input Policy Network To obtain the normalized continuous motion at time t , Here are the policy network parameters; Each component of the dimension lies in [-1,1]; The first M dimensions are mapped to the output setpoints of M controllable generator units: ； Will The N-dimensional mapping is represented by the demand response ratio of N load nodes: ； in, This represents the output setting value of the i-th controllable unit; for The corresponding action component of the i-th controllable unit. for The action component corresponding to the j-th load node, This represents the upper limit of the rated demand response ratio of the j-th load node; M and N are the number of controllable units and the number of load nodes, respectively, and the (M+j)-th dimension represents the j-th dimension in the last N dimensions; and These are the minimum and maximum outputs of the i-th controllable unit, respectively; The demand response ratio of the j-th load node; Let the predictive uncertainty characteristic used by the current scheduling strategy be . The action boundary contraction coefficient is calculated using the following formula: ； in, The coefficient of contraction of the action boundary at time t; Represents the clipping function; To preset the shrinkage gain, ; To preset the minimum action boundary contraction coefficient, ; The lower limit of dynamic output of the i-th controllable unit at time t and dynamic output limit They are respectively: ； ； in, The maximum ramp rate for the i-th controllable unit; The actual output of the i-th controllable unit at time t-1; The upper limit of the dynamic demand response ratio of the j-th load node at time t. for: ； in, This represents the upper limit of the rated demand response ratio for the j-th load node; When motion boundary adjustment is enabled, the rated output upper and lower limits are replaced with dynamic output upper and lower limits, and the normalized continuous motion is mapped to the output setpoint and demand response ratio: ； ； in, This represents the output setpoint of the i-th controllable unit at time t; This represents the demand response ratio of the j-th load node at time t.

9. The microgrid dispatching method based on causal probabilistic wind power forecasting and uncertainty perception according to claim 8, characterized in that, In step S7, the microgrid environment performs constraint updates according to the following expression: ； ； The load regulation constraints are: ； The output of the gas-balanced generator unit is: ； ； in, This represents the intermediate output of the i-th controllable unit at time t; The actual output of the i-th controllable unit at time t. The actual output of the i-th controllable unit at time t-1. This represents the output setpoint of the i-th controllable unit at time t. Let be the maximum ramp rate of the i-th controllable unit. and Let be the minimum and maximum output of the i-th controllable unit, respectively. This represents the upper limit of the rated demand response ratio for the j-th load node; and Let represent the loads before and after adjustment at the j-th load node at time t, respectively. The total load after adjustment at time t; The total controllable unit output at time t; The intermediate output of the gas balance unit at time t; Let t be the actual output of the gas-balanced generator unit. This represents the actual output of the gas-balanced turbine unit at time t-1. This represents the maximum ramp rate for the gas-balanced generator unit. and These are the minimum and maximum outputs of the gas-balanced generator unit, respectively. Let t be the wind power received at time t; Residual supply and demand imbalance at time t for: ； when At that time, reduce the wind power capacity accepted by the following expression: ； ； in, Let t be the wind curtailment power. This indicates that an assignment update is being performed.

10. The microgrid dispatching method based on causal probabilistic wind power prediction and uncertainty perception according to claim 9, characterized in that, In steps S6 and S8, the evaluation metrics are calculated using a technical evaluation function. ； in, This represents the technical evaluation function at time t. , , and To evaluate the weights, and These represent the imbalance penalty coefficients for under-generated and over-generated states, respectively; a continuous-action reinforcement learning agent is used to reduce... The scheduling policy is updated to achieve the target, and the scheduling policy evaluation segment is based on... Measure and evaluate dispatch strategies based on wind power acceptance capacity, wind curtailment capacity, and residual supply and demand imbalance; For the k-th online update cycle, which contains H time points, let's denote the periodic average technical evaluation value when the gating coefficient is not disturbed. And add a preset disturbance to the gating coefficient of the i-th input channel. Post-period average technical evaluation They are respectively: ； in, This represents the technical evaluation function value at time t when the gate coefficients are not disturbed. This indicates that a preset perturbation is applied to the gating coefficient of the i-th input channel at time t. The subsequent technical evaluation function value; Calculate the improvement in evaluation for the i-th input channel based on the periodic average technical evaluation before and after the disturbance: ； in, Let represent the evaluation improvement amount of the i-th input channel in the k-th online update cycle; normalize the evaluation improvement amount of each input channel to obtain the scheduling feedback amount of the i-th input channel: ； in, This represents the scheduling feedback quantity of the i-th input channel in the k-th online update cycle; This represents the minimum improvement in the evaluation of the input channel during the k-th online update cycle. This represents the maximum value of the evaluation improvement of the input channel in the k-th online update cycle; To prevent constants with a denominator of zero; The channel causality weights from the previous online update cycle are weighted and fused with the scheduling feedback, and then the amplitude is limited to obtain the updated channel causality weights: ； in, This represents the channel causality weight of the i-th input channel in the (k+1)-th online update cycle. This represents the channel causality weight of the i-th input channel in the k-th online update cycle; The output bias of the causal channel gating module is recalculated using the updated channel causal weights: ； in, This represents the output bias of the i-th input channel of the causal channel gating module in the (k+1)-th online update cycle. This is the set of times corresponding to the kth online update cycle; This is the preset disturbance amount for the gating coefficient; To preset the online update step size, ; Preset minimum channel causality weight; To prevent constants with a denominator of zero.

Images