New energy output prediction method and system based on Monte Carlo Dropout

Abstract

The invention provides a new energy output prediction method and system based on Monte Carlo Dropout, and the method comprises the steps: carrying out the probabilistic prediction of new energy output through a Monte Carlo Dropout technology, generating a dynamic prediction result containing a confidence interval, and quantifying the uncertainty of meteorological sudden change and equipment state; secondly, constructing a multi-stage random dynamic programming model, discretizing a prediction interval into multi-scene input, designing a non-linear objective function based on the discharge depth, and synchronously optimizing the electricity purchase cost and the energy storage aging cost; and finally, realizing rolling optimization of the system in combination with a model prediction control framework, and dynamically adjusting an energy storage aging cost weight by updating prediction data and a scheduling instruction on line and embedding an energy storage health state real-time feedback mechanism. The photovoltaic and wind power consumption rate can be remarkably improved, the full life cycle cost of an energy storage system is reduced, and meanwhile, the robustness of a scheduling strategy in extreme weather is ensured.

Description

Technical Field

[0001] This application relates to the technical field of image segmentation recognition and analysis, and particularly to a new energy output prediction method and device based on Monte Carlo Dropout. Background Technique

[0002] With the accelerating transformation of the global energy structure towards a clean type, the penetration rate of new energy sources such as wind power and photovoltaic power in the power system continues to increase. However, their output is affected by the coupled influence of meteorological factors and equipment status, showing significant short-term fluctuation characteristics. Therefore, the power system needs to rely on the rapid charge and discharge of energy storage devices to achieve power balance. At the same time, the formulation of dispatching strategies highly depends on the ability of high-precision short-term prediction and risk quantification. However, in the new energy prediction-dispatching system, traditional deterministic prediction models only output single-point estimated values and cannot quantify the probability distribution of prediction errors, resulting in overly conservative or risky dispatching strategies. Secondly, the charge and discharge optimization of energy storage systems is mostly based on fixed aging cost coefficients and fails to dynamically associate with the actual working conditions of the battery (such as SOC, temperature, cycle depth), resulting in an imbalance between life loss and economy. Especially when participating in the power spot market trading, the above defects will amplify the price fluctuation risk and lead to a significant increase in the revenue loss rate of the operator.

[0003] With the accelerating transformation of the global energy structure towards a clean type, the penetration rate of new energy sources such as wind power and photovoltaic power in the power system continues to increase. However, their output is affected by the coupled influence of meteorological factors and equipment status, showing significant short-term fluctuation characteristics. Therefore, the power system needs to rely on the rapid charge and discharge of energy storage devices to achieve power balance. At the same time, the formulation of dispatching strategies highly depends on the ability of high-precision short-term prediction and risk quantification. However, in the new energy prediction-dispatching system, traditional deterministic prediction models only output single-point estimated values and cannot quantify the probability distribution of prediction errors, resulting in overly conservative or risky dispatching strategies. Secondly, the charge and discharge optimization of energy storage systems is mostly based on fixed aging cost coefficients and fails to dynamically associate with the actual working conditions of the battery (such as SOC, temperature, cycle depth), resulting in an imbalance between life loss and economy. Especially when participating in the power spot market trading, the above defects will amplify the price fluctuation risk and lead to a significant increase in the revenue loss rate of the operator.

[0004] In the early stage, the new energy scheduling system adopted a decoupled architecture of "prediction - optimization". For example, deterministic predictions were generated based on algorithms such as ARIMA or SVM, and then energy storage plans were formulated through linear programming. Such methods often led to a mismatch between scheduling instructions and actual output due to ignoring the transmission of prediction uncertainties. In recent years, some studies have introduced stochastic optimization or robust optimization to handle uncertainties, but they rely on preset probability distribution assumptions and are difficult to characterize the true form of prediction errors under complex meteorological conditions. At the same time, deep learning models have improved prediction accuracy, but their black - box characteristics result in insufficient uncertainty quantification and a lack of coordination mechanism with physical constraints. In the energy storage scheduling aspect, existing dynamic programming methods mostly adopt simplified linear aging models, which cannot reflect the non - linear characteristics of battery degradation; although some studies have tried to optimize charge - discharge strategies by combining reinforcement learning, the offline training mode is difficult to adapt to real - time meteorological mutations and does not form a closed - loop feedback with prediction uncertainties.

[0005] Therefore, designing a new energy scheduling method and system that can significantly improve the accommodation rate of photovoltaic and wind power, reduce the full - life - cycle cost of energy storage systems, and ensure the robustness of scheduling strategies under extreme weather conditions in the hybrid operation environment of the power market and new energy micro - grid is an important research content for technical personnel in this field. Summary of the Invention

[0006] To solve the above - mentioned problems, the present invention provides a new energy scheduling method and system based on Monte Carlo Dropout and dynamic programming, aiming to significantly improve the accommodation rate of photovoltaic and wind power, reduce the full - life - cycle cost of energy storage systems, and ensure the robustness of scheduling strategies under extreme weather conditions in the hybrid operation environment of the power market and new energy micro - grid.

[0007] In the first aspect, an embodiment of the present application provides a new energy output prediction method based on Monte Carlo Dropout. Step 1: Collect historical new energy output data, meteorological and environmental data, and real - time data of energy storage systems and market electricity prices;

[0008] Step 2: Pre - process the acquired data;

[0009] Step 3: Perform multiple forward propagations based on a Bi - LSTM model with a Dropout layer to generate a probability distribution of new energy output predictions, extract the quantile interval of the prediction results through the Monte Carlo method, and construct multi - scenario inputs;

[0010] Step 4: Take the minimization of power purchase cost and energy storage aging cost as the optimization objective, and power balance and energy storage capacity as constraints, establish a non - linear dynamic programming model for the new energy - energy storage scheduling system, and use dynamic programming to solve the optimal charge - discharge strategy to achieve cost minimization under risk constraints;

[0011] Step 5: Based on the collaborative cooperation of Monte Carlo prediction and dynamic programming, realize the rolling optimization framework of new energy output prediction and scheduling strategy.

[0012] Optionally, in an implementation manner of the first aspect of the present invention, the step 1: Collect historical output data of new energy, meteorological and environmental data, and real-time data of energy storage systems and market electricity prices, including:

[0013] Data collection consists of three parts: historical output data of new energy, meteorological and environmental data, and energy storage system and market electricity price parameters;

[0014] Among them, the new energy output data refers to the historical multi-scale output data of photovoltaic and wind power with a time span from 15 minutes to 72 hours;

[0015] The meteorological and environmental data includes the meteorological numerical characteristics of wind speed, irradiance, temperature, and cloud cover inherent in the new energy;

[0016] The data of the energy storage and electricity price market are mainly energy storage system parameters, load demand curves, and real-time electricity price data.

[0017] Optionally, in an implementation manner of the first aspect of the present invention, the step 2: Preprocess the obtained data, including:

[0018] Use linear interpolation method to fill short-term missing data and perform outlier removal processing;

[0019] Clean the data, and construct time series features and spatial features for the obtained feature data;

[0020] Perform Min-Max normalization processing on the dynamic features of new energy output data and meteorological information:

[0021]

[0022] Among them, x represents the original data, x norm represents the standard value, x min represents the minimum value of the data set, x max represents the maximum value of the data set.

[0023] Optionally, in an implementation manner of the first aspect of the present invention, the step 3: Perform multiple forward propagations based on the Bi-LSTM model with a Dropout layer, generate the probability distribution of new energy output prediction, and extract the quantile interval of the prediction result through the Monte Carlo method to construct a multi-scenario input, including:

[0024] Use a double-layer bidirectional Bi-LSTM deep neural network as the basic time series prediction model, consider both the forward and backward information of the sequence, and capture the dependencies in the sequence;

[0025] The model inputs temporal features, meteorological features, and time encoding. The input dimension is the number of features × the historical time steps, and the time encoding is represented by a periodic cosine function encoding method;

[0026] Add a Dropout layer after each LSTM layer. Keep it activated during both training and inference, and the dropout probability p = 0.2 - 0.5; During each training, the Dropout layer randomly drops some neurons with a set probability, so that the model cannot overly rely on certain specific neurons, thereby forcing the model to learn more robust feature representations;

[0027] The output layer of the model is a fully connected network, which is responsible for mapping the features extracted by the previous layers to the final prediction result. The output layer is used to predict the predicted values of new energy output points in the next 24 hours At the same time, generate a 90% confidence interval To reflect the uncertainty of the prediction result, where represents the minimum value of the predicted value at time t, represents the maximum value of the predicted value at time t.

[0028] Optionally, in an implementation manner of the first aspect of the present invention, the introduction of the Monte Carlo method to enhance the uncertainty estimation of the Bi-LSTM model includes:

[0029] During the inference stage of the Bi-LSTM model, perform N forward propagations on the same input data. Each time during forward propagation, the Dropout layer randomly drops neurons to generate different prediction results, and the prediction results form a prediction sample set {y1, y2,..., y N};

[0030] The combination of discarded neurons of different neurons will cause the model to generate different predictions. By sampling multiple times to obtain multiple different prediction results, the uncertainty estimation of the model for the input data is captured;

[0031] The Bi-LSTM model processes the input temporal features, meteorological features, and time encoding, and outputs the predicted values of new energy output in the next 24 hours through forward propagation; Calculate the confidence interval based on the prediction results of the Bi-LSTM and the Monte Carlo method to provide probabilistic input for subsequent scheduling optimization, including: Sort the sampling results in ascending order:

[0032] Let y1 ≤ y2 ≤... ≤ y N :

[0033] Determine the quantiles according to the required confidence level. For a 90% confidence interval, calculate the 5% and 95% quantiles as the upper and lower limits of the quantiles:

[0034] Lower_Bound = Q0.05 ({y N )

[0035] Upper_Bound = Q 0.95 ({y N )

[0036] Wherein, Upper_Bound and Lower_Bound are the upper and lower limits of the confidence interval, and Q is the quantile function;

[0037] If the prediction result follows a Gaussian distribution, then further calculate the mean μ and standard deviation σ of the sample set, and generate the interval [μ - kσ, μ + kσ], where the value of k is an adjustment factor determined according to the required confidence level. For a 90% confidence interval, the value of k is approximately 1.645 under the Gaussian distribution;

[0038] If the prediction result does not follow a Gaussian distribution, then continue to use the non-parametric fraction method to generate the confidence interval.

[0039] Optionally, in an implementation manner of the first aspect of the present invention, the step 4: taking the minimization of the electricity purchase cost and the energy storage aging cost as the optimization objective, and taking the power balance and the energy storage capacity as the constraints, establish a non-linear dynamic programming model of the new energy - energy storage scheduling system, and use dynamic programming to solve the optimal charge and discharge strategy to achieve cost minimization under risk constraints, including:

[0040] S31. Construction of the dynamic programming model:

[0041] The definition of the state equation variables of the model is as follows:

[0042] S t =(E t , F t , P t buy , D t )

[0043] Where: E t is the energy storage power at time t, F t is the new energy predicted output interval [F t min , F t max ,

[0044] F t min is the lower limit of the new energy predicted output interval, F t max is the upper limit of the new energy predicted output interval, P t buy is the real-time electricity price, D t is the load demand;

[0045] The decision variable equations are defined as follows:

[0046] The charge and discharge power of the energy storage is u t Satisfy |u t | ≤ u max , charging is positive, discharging is negative, and u max represents the maximum discharging power;

[0047] State transition, that is, the energy storage power update equation is defined as follows:

[0048]

[0049] where η charge and η discharge are the charge / discharge efficiencies;

[0050] The objective function of the model takes the minimum of the system's electricity purchase cost and the energy storage aging cost as the total cost. When the new energy output and the energy storage discharge are not enough to meet the load, the externally purchased electricity is:

[0051]

[0052] where C age is the energy storage aging cost, which reflects the impact of the charge and discharge depth and SOC on the battery life. β is the coefficient that dynamically updates the energy storage health state according to the energy storage cycle aging, and T is the final time period;

[0053]

[0054] where Q total is the initial total capacity, Q loss is the cumulative capacity loss, and β0 is the coefficient of the initial energy storage health state;

[0055] Based on the rainflow counting method, an energy storage aging cost function is established:

[0056]

[0057] where α = 0.002, γ = -0.1;

[0058] The new energy - energy storage scheduling model needs to satisfy power balance and consider the charge and discharge and capacity constraints of the energy storage:

[0059] Power balance: F t + u t + P grid,t = D t , P grid,t ≥ 0 is the externally purchased electricity. Among them, considering the randomness of the new energy output prediction interval: F t ~Uniform(F t min,F t max );

[0060] Energy storage safety capacity constraint: 20% ≤ SOC t ≤ 90%;

[0061] Energy storage charge and discharge power limit: |u t | ≤ u max ;

[0062] S32. Monte Carlo multi-scenario input and dynamic programming solution:

[0063] Obtain the system cost under different scenarios through Monte Carlo simulation, which helps to evaluate the economy and stability of the power optimization dispatch system under different new energy output conditions, including:

[0064] Discretize the prediction interval [F t min ,F t max generated by Monte Carlo into K scenario inputs in dynamic programming, and assign a probability of p k to each scenario. These scenarios represent the output states of new energy under different possibilities. By discretization, the continuous distribution is transformed into multi-scenario inputs and simplified for solution by dynamic programming;

[0065] The new energy output F t in each time period t follows a uniform distribution:

[0066] F t ~Uniform(F t min ,F t max )

[0067] In state S t , select the optimal charge and discharge power u t , and recursively solve the sum of minimizing the current cost and the future expected cost through the Bellman equation:

[0068]

[0069] where: C t is the immediate cost; is the optimal value function in the next time period state;

[0070] By continuously iterating this equation, recursively deduce from the final time period T to the initial time period t = 1. Finally, obtain the optimal decision sequence within the entire scheduling period, calculate the optimal value function and strategy of each state, that is, the optimal value of the charge and discharge power u t of the energy storage at different times, to minimize the total cost.

[0071] Optionally, in an implementation of the first aspect of the present invention, step 5: Based on the collaborative cooperation of Monte Carlo prediction and dynamic programming, an rolling optimization framework for new energy output prediction and scheduling strategy is implemented, including:

[0072] Combined with the model predictive control framework, every 15 minutes, a confidence interval is regenerated based on the Bi-LSTM model and the Monte Carlo Dropout method, and the scheduling strategy for the next few hours is dynamically updated based on the dynamic programming algorithm to generate the latest energy storage scheduling instruction, in response to the real-time change of data, including:

[0073] S41. Optimize the energy storage scheduling strategy within the optimization window and execute the optimal action u at the current moment t * ;

[0074] S42. Collect the actual output data of new energy and the energy storage state, and correct the confidence interval of the new energy output generated by the prediction model;

[0075] S43. Collect the actual energy storage state and new energy output data, and update the energy storage aging cost coefficient β;

[0076] The Bi-LSTM model combines the Monte Carlo Dropout method to generate the confidence interval of the new energy output as the predicted-to-optimized multi-scenario input of the dynamic programming. At the same time, the actual energy storage state data and new energy output data are fed back to the next round of optimization-to-prediction, so as to correct the deviation between the model prediction and the scheduling.

[0077] In the second aspect, an embodiment of the present application provides a new energy output prediction system based on Monte Carlo Dropout, which is applied to the new energy output prediction method based on Monte Carlo Dropout as described in the first aspect, and is characterized by including:

[0078] Collection module: Collect the historical output data of new energy, meteorological and environmental data, and the real-time data of the energy storage system and market electricity price;

[0079] Preprocessing module: Preprocess the acquired data;

[0080] Prediction module: Based on the Bi-LSTM model with a Dropout layer, perform multiple forward propagations to generate the probability distribution of the new energy output prediction, extract the quantile interval of the prediction result through the Monte Carlo method, and construct a multi-scenario input;

[0081] Solving module: Taking the minimization of the electricity purchase cost and the energy storage aging cost as the optimization goal, and taking the power balance and energy storage capacity as the constraints, establish a non-linear dynamic programming model for the new energy-storage scheduling system, and use dynamic programming to solve the optimal charge and discharge strategy to achieve the cost minimization under risk constraints;

[0082] Execution module: Based on the collaborative cooperation of Monte Carlo prediction and dynamic programming, a rolling optimization framework for new energy output prediction and scheduling strategy is realized.

[0083] Thirdly, an embodiment of the present application provides an electronic device, including:

[0084] A processor;

[0085] A memory for storing instructions executable by the processor;

[0086] Wherein, when the processor is configured to execute the instructions, the new energy output prediction method based on Monte Carlo Dropout as described in the first aspect is implemented.

[0087] Fourthly, an embodiment of the present application provides a computer-readable storage medium, and the computer-readable storage medium stores a program, and the program instructs the device to execute the new energy output prediction method based on Monte Carlo Dropout as described in the first aspect.

[0088] In the technical solution provided by the present invention, a new energy output prediction method and device based on Monte Carlo Dropout are provided. Through the Monte Carlo Dropout technology, the new energy output is probabilistically predicted to generate a dynamic prediction result including a confidence interval, and the uncertainties of meteorological mutations and equipment states are quantified. Secondly, a multi-stage stochastic dynamic programming model is constructed, the prediction interval is discretized into multi-scenario inputs, and a non-linear objective function based on the depth of discharge is designed to synchronously optimize the electricity purchase cost and the energy storage aging cost. Finally, combined with the model predictive control framework, the rolling optimization of the system is realized. By online updating the prediction data and scheduling instructions and embedding a real-time feedback mechanism for the health state of the energy storage, the weight of the energy storage aging cost is dynamically adjusted. It can significantly improve the consumption rate of photovoltaic and wind power, reduce the full life cycle cost of the energy storage system, and at the same time ensure the robustness of the scheduling strategy in extreme weather.

[0089] Beneficial effects:

[0090] 1. Through the coupling mechanism of Monte Carlo Dropout and dynamic programming, a confidence interval for new energy output is generated and discretized into multi-scenario inputs, quantifying the impact of uncertainties on scheduling, enhancing the decision-making robustness in extreme weather, and reducing the revenue loss caused by prediction deviation;

[0091] 2. Based on the dynamic embedding of the charge and discharge depth and non-linear aging cost, accurately reflecting the attenuation characteristics of the energy storage, extending the life of the energy storage while reducing the operation and maintenance cost;

[0092] 3. Design a "prediction-optimization" closed-loop system that dynamically updates interval predictions and scheduling strategies every 15 minutes, and achieves second-level response with the support of hardware resources to ensure that the power system meets real-time power balance and improves arbitrage benefits in the scenario of power market fluctuations. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 It is a schematic flowchart of a new energy output prediction method based on Monte Carlo Dropout provided by an embodiment of the present application.

[0094] Figure 2 It is a schematic diagram of a new energy scheduling method provided by an embodiment of the present application.

[0095] Figure 3 It is a schematic diagram of a new energy output prediction system module based on Monte Carlo Dropout provided by an embodiment of the present application.

[0096] Figure 4 It is a schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0097] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments.

[0098] It should be noted that "at least one" in the embodiments of the present application means one or more, and multiple means two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application.

[0099] It should be noted that in the embodiments of the present application, terms such as "first" and "second" are only used for the purpose of distinguishing descriptions, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying order. Features defined with "first" and "second" can explicitly or implicitly include one or more of the described features. In the description of the embodiments of the present application, words such as "exemplary" or "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, using words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0100] Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of this application.

[0101] Embodiment 1

[0102] Figure 1 It is a schematic flow diagram of a new energy output prediction method based on Monte Carlo Dropout provided by an embodiment of this application.

[0103] As Figure 1 shown, this application provides a new energy output prediction method based on Monte Carlo Dropout, including the following steps.

[0104] Step 1: Collect historical new energy output data, meteorological and environmental data, and real-time data of energy storage systems and market electricity prices.

[0105] Specifically, data collection mainly consists of three parts: historical new energy output data, meteorological and environmental data, and energy storage system and market electricity price parameters. Among them, new energy output data refers to historical multi-scale output data of photovoltaic and wind power with a time span from 15 minutes to 72 hours; meteorological environment data includes meteorological numerical characteristics such as wind speed, irradiance, temperature, cloud cover, etc. inherent in the new energy; and data of energy storage and electricity price markets are mainly energy storage system parameters (capacity, efficiency, SOC limit), load demand curves, real-time electricity price data, etc.

[0106] Step 2: Preprocess the obtained data.

[0107] Specifically, the step 2: preprocess the obtained data, including:

[0108] Use linear interpolation method to fill short-term missing data and perform outlier rejection processing;

[0109] Clean the data, and construct time series features and spatial features for the obtained feature data;

[0110] Perform Min-Max normalization processing on the dynamic features of new energy output data and meteorological information:

[0111]

[0112] Among them, x represents the original data, x norm represents the standard value, x min represents the minimum value of the data set, x max represents the maximum value of the data set.

[0113] Specifically, the acquired data is preprocessed, such as filling short-term missing data (≤3 time points) using linear interpolation method and removing outliers. The cleaned feature data is used to construct time series features (lag variables, sliding window statistics (24-hour sliding mean, variance)) and spatial features (constructing an adjacent power station output correlation matrix based on Pearson correlation coefficient (>0.8)).

[0114] Step 3: Based on the Bi-LSTM model with Dropout layer, perform multiple forward propagations to generate the probability distribution of new energy output prediction. Extract the quantile interval of the prediction result through the Monte Carlo method to construct a multi-scenario input.

[0115] Specifically, Figure 2 This is a schematic diagram of a new energy scheduling method provided by an embodiment of the present application. As Figure 2 shown, the step 2: Based on the Bi-LSTM model with Dropout layer, perform multiple forward propagations to generate the probability distribution of new energy output prediction. Extract the quantile interval of the prediction result through the Monte Carlo method to construct a multi-scenario input, including:[[]]

[0116] Adopt a double-layer bidirectional Bi-LSTM deep neural network as the basic time series prediction model, considering both the forward and backward information of the sequence to capture the dependencies in the sequence;

[0117] The model inputs time series features, meteorological features, and time encoding. The input dimension is the number of features × historical time steps, and the time encoding is represented by a periodic cosine function encoding method;

[0118] Add a Dropout layer after each layer of LSTM, keep it active during both training and inference, and the dropout probability p = 0.2 - 0.5; during each training, the Dropout layer will randomly discard some neurons with a set probability, so that the model cannot overly rely on certain specific neurons, thereby forcing the model to learn more robust feature representations;

[0119] The output layer of the model is a fully connected network, responsible for mapping the features extracted by the previous layers to the final prediction result. The output layer is used to predict the predicted values of new energy output points in the next 24 hours while generating a 90% confidence interval to reflect the uncertainty of the prediction result, where represents the lowest value of the predicted value at time t, represents the highest value of the predicted value at time t.

[0120] Specifically, a double-layer bidirectional Bi-LSTM deep neural network is used as the basic time series prediction model, considering both the forward and backward information of the sequence to capture the dependencies in the sequence more comprehensively. The model inputs time series features, meteorological features, and time encoding (time is represented by the periodic cosine function encoding method), and the input dimension is the number of features × the historical time step.

[0121] To enhance the generalization ability of the model, a Dropout layer is added after each layer of LSTM, which remains active during both training and inference, and the dropout probability p = 0.2 - 0.5. That is, during each training, the Dropout layer randomly discards some neurons with a set probability, so that the model cannot overly rely on certain specific neurons, thus forcing the model to learn more robust feature representations.

[0122] The output layer of the model is a fully connected network, which is responsible for mapping the features extracted by the previous layers to the final prediction result. The output layer not only predicts the predicted values of the new energy output points in the next 24 hours but also generates a 90% confidence interval to reflect the uncertainty of the prediction result.

[0123] Specifically, the Monte Carlo method is introduced to enhance the uncertainty estimation of the Bi-LSTM model, including:

[0124] During the inference stage of the Bi-LSTM model, N forward propagations are performed on the same input data. During each forward propagation, the Dropout layer randomly discards neurons to generate different prediction results, and the prediction results form a prediction sample set {y1, y2,..., y N};

[0125] Different combinations of discarded neurons will cause the model to generate different predictions. By sampling multiple times to obtain multiple different prediction results, the uncertainty estimation of the model for the input data is captured;

[0126] The Bi-LSTM model processes the input time series features, meteorological features, and time encoding, and outputs the predicted values of the new energy output in the next 24 hours through forward propagation; based on the prediction results of the Bi-LSTM and the Monte Carlo method, the confidence interval is calculated to provide probabilistic input for subsequent scheduling optimization, including: sorting the sampling results in ascending order:

[0127] Let y1 ≤ y2 ≤... ≤ y N :

[0128] Determine the quantiles according to the required confidence level. For a 90% confidence interval, calculate the 5% and 95% quantiles as the upper and lower limits of the quantiles:

[0129] Lower_Bound = Q 0.05({y N )

[0130] Upper_Bound = Q 0.95 ({y N )

[0131] Wherein, Upper_Bound and Lower_Bound are the upper and lower limits of the confidence interval, and Q is the quantile function;

[0132] If the prediction result follows a Gaussian distribution, then further calculate the mean μ and standard deviation σ of the sample set, and generate the interval [μ - kσ, μ + kσ]. Among them, the k value is the adjustment factor, which is determined according to the required confidence level. For a 90% confidence interval, under the Gaussian distribution, the k value is approximately 1.645;

[0133] If the prediction result does not follow a Gaussian distribution, then continue to use the non-parametric fractional method to generate the confidence interval.

[0134] Specifically, the present invention introduces the Monte Carlo method to enhance the uncertainty estimation of the Bi-LSTM model, and approximately solves the problem through a large number of random samplings. The inference stage of the Bi-LSTM model is similar to Monte Carlo sampling. For the same input data, perform N forward propagations. Each time during the forward propagation, the Dropout layer randomly discards neurons, thereby generating different prediction results. These results constitute the prediction sample set {y1, y2,..., y N}). Because different combinations of discarded neurons will cause the model to generate different predictions, multiple different prediction results are obtained through multiple samplings, thereby capturing the uncertainty estimation of the model for the input data.

[0135] The Bi-LSTM model processes the input time series features, meteorological features, and time encoding, and outputs the predicted value of the new energy output for the next 24 hours through forward propagation. Calculate the confidence interval based on the prediction result of the Bi-LSTM and the Monte Carlo method, providing a probabilistic input for subsequent scheduling optimization.

[0136] Specifically, sort the sampling results in ascending order:

[0137] Let y1 ≤ y2 ≤... ≤ y N :

[0138] Determine the quantiles according to the required confidence level. For a 90% confidence interval, then calculate the 5% and 95% quantiles as the upper and lower limits of the quantiles:

[0139] Lower_Bound = Q 0.05 ({y N )

[0140] Upper_Bound = Q0.95 ({y N )

[0141] Among them, Upper_Bound and Lower_Bound are the upper and lower limits of the confidence interval, and Q is the quantile function.

[0142] If the prediction result follows a Gaussian distribution, further calculate the mean μ and standard deviation σ of the sample set, and generate the interval [μ - kσ, μ + kσ]. Among them, the value of k is determined according to the required confidence level. For a 90% confidence interval, the value of k is approximately 1.645 under the Gaussian distribution; if the prediction result does not follow a Gaussian distribution, then continue to use the above non-parametric fractional method to generate the confidence interval.

[0143] Step 4: Taking the minimization of the electricity purchase cost and the energy storage aging cost as the optimization objective, and taking the power balance and the energy storage capacity as the constraints, establish a non-linear dynamic programming model for the new energy - energy storage scheduling system, and use dynamic programming to solve the optimal charge and discharge strategy to achieve the cost minimization under risk constraints.

[0144] Specifically, the said Step 4: Taking the minimization of the electricity purchase cost and the energy storage aging cost as the optimization objective, and taking the power balance and the energy storage capacity as the constraints, establish a non-linear dynamic programming model for the new energy - energy storage scheduling system, and use dynamic programming to solve the optimal charge and discharge strategy to achieve the cost minimization under risk constraints, includes:

[0145] S31. Construction of the dynamic programming model:

[0146] The definition of the state equation variables of the model is as follows:

[0147] S t =(E t , F t , P t buy , D t )

[0148] Among them: E t is the energy storage power at time t, F t is the predicted new energy output interval [F t min , F t max ,

[0149] F t min is the lower limit of the predicted new energy output interval, F t max is the upper limit of the predicted new energy output interval, P t buy is the real-time electricity price, D t is the load demand;

[0150] The decision variable equations are defined as follows:

[0151] The charge and discharge power of the energy storage, u t satisfies |u t | ≤ u max , where charging is positive and discharging is negative, and u max represents the maximum discharging power;

[0152] State transition, that is, the energy storage state update equation is defined as follows:

[0153]

[0154] where η charge and η discharge are the charge / discharge efficiencies;

[0155] The objective function of the model takes the minimum of the system's electricity purchase cost and the energy storage aging cost as the total cost. When the new energy output and the energy storage discharge are not enough to meet the load, the purchased electricity quantity:

[0156]

[0157] where C age is the energy storage aging cost, which reflects the impact of the charge and discharge depth and the SOC on the battery life. β is the coefficient that dynamically updates the energy storage health state according to the energy storage cycle aging, and T is the final time period;

[0158]

[0159] where Q total is the initial total capacity, Q loss is the cumulative capacity loss, and β0 is the coefficient of the initial energy storage health state;

[0160] Based on the rainflow counting method, an energy storage aging cost function is established:

[0161]

[0162] where α = 0.002 and γ = -0.1;

[0163] The new energy - energy storage scheduling model needs to satisfy the power balance and consider the charge and discharge and capacity constraints of the energy storage:

[0164] Power balance: F t + u t + P grid,t = D t , P grid,t ≥ 0 is the purchased electricity quantity. Among them, considering the randomness of the new energy output prediction interval: F t ~ Uniform(F t min,F t max );

[0165] Energy storage safety capacity constraint: 20% ≤ SOC t ≤ 90%;

[0166] Energy storage charge and discharge power limit: |u t | ≤ u max .

[0167] S32. Monte Carlo multi-scenario input and dynamic programming solution:

[0168] Obtaining the cost situation of the system under different scenarios through Monte Carlo simulation helps to evaluate the economy and stability of the power optimization scheduling system under different new energy output conditions, including:

[0169] Discretize the prediction interval [F t min ,F t max generated by Monte Carlo into K scenario inputs in dynamic programming, and the probability assigned to each scenario is p k , and these scenarios represent the output states of new energy under different possibilities. By discretization, the continuous distribution is transformed into multi-scenario inputs and simplified to dynamic programming for solution;

[0170] The new energy output F t in each time period t follows a uniform distribution:

[0171] F t ~Uniform(F t min ,F t max )

[0172] In state S t , select the optimal charge and discharge power u t , and recursively solve the sum of minimizing the current cost and the future expected cost through the Bellman equation:

[0173]

[0174] where: C t is the immediate cost; is the optimal value function in the next time period state;

[0175] By continuously iterating this equation, recursively deduce from the final time period T to the initial time period t = 1, and finally obtain the optimal decision sequence within the entire scheduling period, calculate the optimal value function and strategy of each state, that is, the optimal value of the charge and discharge power u t of the energy storage at different times, to minimize the total cost.

[0176] Step 5: Based on the collaborative cooperation of Monte Carlo prediction and dynamic programming, implement a rolling optimization framework for new energy output prediction and scheduling strategies.

[0177] Specifically, the Step 5: Based on the collaborative cooperation of Monte Carlo prediction and dynamic programming, implement a rolling optimization framework for new energy output prediction and scheduling strategies, including:

[0178] Combined with the model predictive control framework, regenerate the confidence interval based on the Bi-LSTM model and the Monte Carlo Dropout method every 15 minutes, and dynamically update the scheduling strategy for the next few hours based on the dynamic programming algorithm to generate the latest energy storage scheduling instructions, responding to the real-time changes in data, including:

[0179] S41. Optimize the energy storage scheduling strategy within the optimization window and execute the optimal action at the current moment

[0180] S42. Collect the actual new energy output data and the energy storage state, and correct the confidence interval of the new energy output generated by the prediction model;

[0181] S43. Collect the actual energy storage state and the new energy output data, and update the energy storage aging cost coefficient β;

[0182] The Bi-LSTM model combines the Monte Carlo Dropout method to generate the confidence interval of the new energy output as the multi-scenario input from prediction to optimization of dynamic programming. At the same time, the actual energy storage state data and the new energy output data are fed back to the next round of optimization to prediction, so as to correct the deviation between model prediction and scheduling.

[0183] Specifically, the Bi-LSTM model combines the Monte Carlo Dropout method to generate the confidence interval of the new energy output as the multi-scenario input (from prediction to optimization) of dynamic programming. At the same time, the actual energy storage state data and the new energy output data are fed back to the next round of prediction (from optimization to prediction), so as to correct the deviation between model prediction and scheduling.

[0184] Embodiment 2

[0185] As Figure 3 shown, the present application provides a new energy output prediction system based on Monte Carlo Dropout, which is applied to the new energy output prediction method based on Monte Carlo Dropout as described in Embodiment 1, including: a collection module 11, a preprocessing module 12, a prediction module 13, a solution module 14, and an execution module 15.

[0186] It can be understood that, in this embodiment, the collection module 11: collects the historical new energy output data, meteorological and environmental data, and the real-time data of the energy storage system and the market electricity price;

[0187] It can be understood that in this embodiment, the preprocessing module 12 preprocesses the acquired data.

[0188] It can be understood that in this embodiment, the prediction module 13 performs multiple forward propagations based on a Bi-LSTM model with a Dropout layer, generates the probability distribution of new energy output prediction, extracts the quantile interval of the prediction result through the Monte Carlo method, and constructs a multi-scenario input.

[0189] It can be understood that in this embodiment, the solving module 14 takes the minimization of the power purchase cost and the energy storage aging cost as the optimization objective, takes power balance and energy storage capacity as constraints, establishes a non-linear dynamic programming model for the new energy-storage scheduling system, and uses dynamic programming to solve the optimal charge and discharge strategy to achieve cost minimization under risk constraints.

[0190] It can be understood that in this embodiment, the execution module 15 realizes the rolling optimization framework of new energy output prediction and scheduling strategy based on the collaborative cooperation of Monte Carlo prediction and dynamic programming.

[0191] Figure 4 This is an electronic device provided by an embodiment of the present application. As Figure 4 shown, the electronic device at least includes the following parts: a processor 101, a memory 100, a communication interface 103, and a bus 102.

[0192] In the embodiment of the present application, the memory 100 is used to store executable instructions of the processor 101, and the processor 101 is configured to execute the instructions to implement the device modules as Figure 3 shown.

[0193] In the embodiment of the present application, a computer-readable storage medium includes instructions that direct the device to execute the method in the first aspect. For example, the instructions direct the device to execute the process steps as Figure 1 shown.

[0194] The program operating in the electronic device involved in an embodiment of the present application can be a program that controls a central processing unit (CPU) and the like to implement the functions of the above-mentioned implementation manners involved in a solution of the present invention (a program that enables a computer to function). Then, the information processed by these devices is temporarily stored in a random access memory (RAM) during its processing, and then stored in various ROMs such as a read-only memory (Flash ROM), a hard disk drive (HDD), etc., and read, corrected, and written by the CPU as needed.

[0195] It should be noted that part of the electronic device according to the above-described embodiments can also be implemented by a computer. In this case, a program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into a computer and executed to implement it.

[0196] It should be noted that the "computer" mentioned here refers to a computer built into an electronic device, which is a computer including hardware such as an OS and peripheral devices. In addition, the "computer-readable recording medium" refers to removable media such as a floppy disk, magneto-optical disk, ROM, CD-ROM, and storage devices such as a hard disk built into a computer.

[0197] Moreover, the "computer-readable recording medium" can include: a medium that stores a program dynamically for a short period of time, such as a communication line when a program is transmitted via a network such as the Internet or a communication line such as a telephone line; a medium that stores a program for a fixed period of time, such as a volatile memory inside a computer of a server or a client in this case. In addition, the above program can be a program for implementing a part of the above functions, and can also be a program that can implement the above functions by combining with a program already recorded in a computer.

[0198] In addition, the electronic device in the above-described embodiments can also be implemented as an aggregate (device group) composed of multiple devices. Each device constituting the device group can have some or all of the functions or function blocks of the electronic device according to the above-described embodiments. As the device group, it is sufficient to have all the functions or function blocks of the electronic device.

[0199] Those of ordinary skill in the art in this technical field should recognize that the above embodiments are only used to illustrate the present application, rather than to limit the present application. As long as it is within the scope of the substantial spirit of the present application, appropriate changes and variations made to the above embodiments fall within the scope of protection required by the present application.

Claims

1. A new energy output prediction method based on Monte Carlo Dropout, characterized in that The method includes: Step 1: Collect historical output data of new energy, meteorological and environmental data, and real-time data of energy storage systems and market electricity prices; Step 2: Preprocess the acquired data; Step 3: Perform multiple forward propagations based on a Bi-LSTM model with a Dropout layer to generate a probability distribution of new energy output prediction, extract the quantile interval of the prediction result through the Monte Carlo method, and construct multi-scenario inputs; Step 4: With the minimization of power purchase cost and energy storage aging cost as the optimization objective, and power balance and energy storage capacity as constraints, establish a non-linear dynamic programming model for the new energy-storage scheduling system, and use dynamic programming to solve the optimal charge-discharge strategy to achieve cost minimization under risk constraints; Step 5: Based on the collaborative cooperation of Monte Carlo prediction and dynamic programming, realize a rolling optimization framework for new energy output prediction and scheduling strategy.

2. The new energy output prediction method based on Monte Carlo Dropout according to claim 1, characterized in that The said Step 1: Collect historical output data of new energy, meteorological and environmental data, and real-time data of energy storage systems and market electricity prices, including: Data collection consists of three parts: historical output data of new energy, meteorological environment data, and parameters of energy storage systems and market electricity prices; Among them, the new energy output data refers to historical multi-scale output data of photovoltaic and wind power with a time span from 15 minutes to 72 hours; The meteorological environment data includes the inherent meteorological numerical characteristics of new energy such as wind speed, irradiance, temperature, and cloud cover; The data of the energy storage and electricity price market are mainly energy storage system parameters, load demand curves, and real-time electricity price data.

3. The new energy output prediction method based on Monte Carlo Dropout according to claim 2, wherein The said Step 2: Preprocess the acquired data, including: Use linear interpolation method to fill in short-term missing data and perform outlier removal; Clean the data, and construct time series features and spatial features for the obtained feature data; Perform Min-Max standardization processing on the dynamic characteristics of new energy output data and meteorological information; Among them, x represents the original data, x norm represents the standard value, x min represents the minimum value of the data set, x max represents the maximum value of the data set.

4. The new energy output prediction method based on Monte Carlo Dropout according to claim 1, characterized in that The said Step 3: Perform multiple forward propagations based on a Bi-LSTM model with a Dropout layer to generate a probability distribution of new energy output prediction, extract the quantile interval of the prediction result through the Monte Carlo method, and construct multi-scenario inputs, including: Adopt a double-layer bidirectional Bi-LSTM deep neural network as the basic time series prediction model, consider both the forward and backward information of the sequence, and capture the dependencies in the sequence; The model inputs time series features, meteorological features, and time encoding, and the input dimension is the number of features × historical time steps. The time encoding is represented by a periodic cosine function encoding method; Add a Dropout layer after each layer of LSTM, keep it activated during training and inference, and the dropout probability p = 0.2 - 0.5; during each training, the Dropout layer will randomly discard some neurons with a set probability, so that the model cannot overly rely on certain specific neurons, thereby forcing the model to learn more robust feature representations; The output layer of the model is a fully connected network, which is responsible for mapping the features extracted by the previous layers to the final prediction result. The output layer is used to predict the predicted value of the new energy output point for the next 24 hours. At the same time, a 90% confidence interval is generated. To reflect the uncertainty of the prediction result, where represents the lowest value of the predicted value at time t, represents the highest value of the predicted value at time t.

5. The new energy output prediction method based on Monte Carlo Dropout according to claim 4, characterized in that, The introduction of the Monte Carlo method to enhance the uncertainty estimation of the Bi-LSTM model includes: During the inference stage of the Bi-LSTM model, the same input data is propagated forward N times. Each time during the forward propagation, the Dropout layer randomly discards neurons to generate different prediction results, and the said prediction results form a prediction sample set {y1, y2,..., y N}; Different combinations of neuron discarding will cause the model to generate different predictions. By sampling multiple times to obtain multiple different prediction results, capture the uncertainty estimation of the model for the input data; The Bi-LSTM model processes the input time series features, meteorological features, and time encoding, and outputs the predicted values of new energy output for the next 24 hours through forward propagation; based on the prediction results of Bi-LSTM and the Monte Carlo method, the confidence interval is calculated to provide probabilistic input for subsequent scheduling optimization, including: sorting the sampling results in ascending order: Let y1 ≤ y2 ≤... ≤ y N : Determine the quantiles according to the required confidence level. For the 90% confidence interval, calculate the 5% and 95% quantiles as the upper and lower limits of the quantiles: Lower_Bound=Q 0.05 ({y N}) Upper_Bound=Q 0.95 ({y N}) where Upper_Bound and Lower_Bound are the upper and lower limits of the confidence interval, and Q is the quantile function; If the prediction results follow a Gaussian distribution, further calculate the mean μ and standard deviation σ of the sample set, and generate the interval [μ - kσ, μ + kσ], where the k value is an adjustment factor determined according to the required confidence level. For the 90% confidence interval, the k value is approximately 1.645 under the Gaussian distribution; If the prediction results do not follow a Gaussian distribution, continue to use the non-parametric fractional method to generate the confidence interval.

6. The new energy output prediction method based on Monte Carlo Dropout according to claim 5, wherein, Step 4: Taking the minimization of the electricity purchase cost and the energy storage aging cost as the optimization objective, and taking power balance and energy storage capacity as constraints, establish a non-linear dynamic programming model for the new energy-storage scheduling system, and use dynamic programming to solve the optimal charge and discharge strategy to achieve cost minimization under risk constraints, including: S 31. Construction of the dynamic programming model: The definitions of the state equation variables of the model are as follows: S t = (E t , F t , P t buy , D t ) Where: E t is the energy storage power at time t, F t is the predicted output range of new energy [F t min , F t max , F t min is the lower limit of the predicted output range of new energy, F t max is the upper limit of the predicted output range of new energy, P t buy is the real-time electricity price, D t is the load demand; The definitions of the decision variable equations are as follows: Energy storage charge and discharge power u t Satisfy |u t | ≤ u max , charging is positive, discharging is negative, u max represents the maximum discharge power; State transition, that is, the energy storage power update equation is defined as follows: Among them, η charge and η discharge are the charge / discharge efficiency; The objective function of the model takes the minimization of the system electricity purchase cost and the energy storage aging cost as the total cost. When the new energy output and the energy storage discharge are not enough to meet the load, the purchased electricity: Among them, C age is the energy storage aging cost, reflecting the impact of charge-discharge depth and SOC on battery life. β is the coefficient for dynamically updating the energy storage health state according to energy storage cycle aging, and T is the final time period; Among them, Q total is the initial total capacity, Q loss is the cumulative capacity loss, and β0 is the coefficient of the initial energy storage health state; Based on the rain flow counting method, establish an energy storage aging cost function: where α = 0.002 and γ = -0.1; The new energy-storage scheduling model needs to satisfy power balance and consider the energy storage charge and discharge and capacity constraint conditions: Power balance: F t + u t + P grid,t = D t , P grid,t ≥ 0 is the purchased electricity. Among them, considering the randomness of the new - energy output prediction interval: F t ~ Uniform(F t min , F t max ) Energy storage safety capacity constraint: 20% ≤ SOC t ≤ 90%; Energy storage charge and discharge power limit: |u t | ≤ u max ; S 32. Monte Carlo multi-scenario input and dynamic programming solution: Obtain the cost situation of the system under different scenarios through Monte Carlo simulation, which helps to evaluate the economy and stability of the power optimization scheduling system under different new energy output conditions, including: The prediction interval [F generated by Monte Carlo t min , F t max is discretized into K scenario inputs in dynamic programming, and the probability assigned to each scenario is p k , and these scenarios represent the output states of new energy under different possibilities. By discretization, the continuous distribution is transformed into multi-scenario inputs, which are simplified to dynamic programming for solution; New energy output F for each period t t Subject to a uniform distribution: F t ~Uniform(F t min ,F t max ) At state S t select the optimal charging and discharging power u t , and recursively solve the Bellman equation to minimize the sum of the current cost and the future expected cost: Where: C t is the immediate cost; is the optimal value function in the state of the next time period; By continuously iterating this equation, we recursively calculate from the final time period T back to the initial time period t = 1, and finally obtain the optimal decision sequence within the entire scheduling period, calculating the optimal value function and policy of each state, that is, the charging and discharging power u of the energy storage at different times t to minimize the total cost.

7. The new energy output prediction method based on Monte Carlo Dropout according to claim 6, characterized in that Step 5: Based on the collaborative cooperation of Monte Carlo prediction and dynamic programming, realize the rolling optimization framework of new energy output prediction and scheduling strategy, including: Combined with the model predictive control framework, regenerate the confidence interval every 15 minutes based on the Bi-LSTM model and the Monte Carlo Dropout method, and dynamically update the scheduling strategy for the next few hours based on the dynamic programming algorithm to generate the latest energy storage scheduling instructions to respond to real-time changes in data, including: S 41. Optimize the energy storage scheduling strategy within the window and execute the optimal action at the current moment S 42. Collect the actual new energy output data and the energy storage state, and correct the confidence interval of the new energy output generated by the prediction model; S 43. Collect the actual energy storage state and the new energy output data, and update the energy storage aging cost coefficient β; The Bi-LSTM model combined with the Monte Carlo Dropout method generates the confidence interval of new energy output as the predicted multi-scenario input for dynamic programming. At the same time, the actual energy storage state data and new energy output data are fed back to the next round of optimization to prediction, thereby correcting the deviation between model prediction and scheduling.

8. A new energy output prediction system based on Monte Carlo Dropout, which is applied to the new energy output prediction method based on Monte Carlo Dropout according to any one of claims 1 to 7, and is characterized in that, It includes: Collection module: Collect historical new energy output data, meteorological and environmental data, as well as real-time data of energy storage systems and market electricity prices; Preprocessing module: Preprocess the acquired data; Prediction module: Based on the Bi-LSTM model with a Dropout layer, perform multiple forward propagations to generate the probability distribution of new energy output prediction. Extract the quantile interval of the prediction result through the Monte Carlo method to construct a multi-scenario input; Solution module: With the minimization of power purchase cost and energy storage aging cost as the optimization objective, and power balance and energy storage capacity as constraints, establish a non-linear dynamic programming model for the new energy-storage scheduling system, and use dynamic programming to solve the optimal charge and discharge strategy to achieve cost minimization under risk constraints; Execution module: Based on the collaborative cooperation of Monte Carlo prediction and dynamic programming, implement a rolling optimization framework for new energy output prediction and scheduling strategy.

9. An electronic device, characterized in that, It includes: Processor; Memory for storing processor-executable instructions; Wherein, the processor is configured to implement the Monte Carlo Dropout-based new energy output prediction method according to any one of claims 1 to 7 when executing the instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program, and the program instructs the device to execute the Monte Carlo Dropout-based new energy output prediction method according to any one of claims 1 to 7.

Images