Power grid dispatching optimization method for predicting power grid load change based on AI algorithm

Abstract

The invention discloses a power grid dispatching optimization method for predicting power grid load change based on an AI algorithm. The method comprises six steps of data preparation, feature engineering, AI model prediction based on quantile regression, robust optimization dispatching based on prediction results, rolling updating and closed-loop feedback. According to the method, an end-to-end hybrid prediction model is adopted, quantile regression is innovatively introduced, a prediction interval is output, and the prediction uncertainty is quantized; according to the method, a prediction interval is directly converted into a robust optimization constraint, a system is forcibly required to reserve up-regulation reserve and down-regulation reserve to cover the interval, and a scheduling scheme with the optimal cost can be found in a known uncertainty range; the method has the advantages of being high in prediction precision, high in uncertainty coping capacity, low in total cost and high in practicability, and power grid dispatching can be changed from empirical decision making based on certainty to intelligent decision making based on probability and data driving.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power grid dispatching optimization, and particularly relates to a power grid dispatching optimization method based on AI algorithm for predicting power grid load change. BACKGROUND

[0002] With the deepening of new energy transformation, power grid dispatching optimization technology is deeply integrated with big data, artificial intelligence, power electronics and other technologies, and develops towards a more intelligent, adaptive and decentralized direction.

[0003] However, the existing power grid dispatching optimization technology has the following problems:

[0004] 1. The plan is based on a single load point prediction value, and once the prediction deviation is large, it may lead to power imbalance, and expensive real-time adjustment is required, and even power outage risk may be caused;

[0005] 2. The most economic dispatching scheme may not have enough backup, and is very fragile, while the safest dispatching scheme reserves too much backup, and the cost is high;

[0006] 3. The randomness and volatility of wind power, photovoltaic and other new energy sources are the main sources of power grid instability, and sometimes the dispatching center has to give up green and environmentally friendly wind power and photovoltaic power for safety;

[0007] In order to solve the above problems, it is necessary to develop a power grid dispatching optimization method based on AI algorithm for predicting power grid load change. SUMMARY

[0008] The purpose of the present application is to overcome the shortcomings of the prior art, and to provide a power grid dispatching optimization method based on AI algorithm for predicting power grid load change, which has high prediction accuracy, strong ability to cope with uncertainty, low total cost and strong practicality, can realize the transition of power grid dispatching from deterministic experience-based decision-making to probability-based and data-driven intelligent decision-making, and solves the economic and safety balance problem of power grid dispatching under high proportion of renewable energy access.

[0009] The purpose of the present application is achieved by a power grid dispatching optimization method based on AI algorithm for predicting power grid load change, comprising the following steps:

[0010] S1, data preparation: collecting every hour dispatching related original data in a certain time range, including load data, weather data, date data and power grid data;

[0011] S2, feature engineering: through extraction, calculation and combination, the collected dispatching related original data is converted into key signals and indicators that can be understood by AI model, including time features, weather features and historical load features;

[0012] S3, AI model prediction based on quantile regression: build an end-to-end hybrid prediction model to accurately predict the point value Q50 of the future grid net load and its fluctuation interval at one time ;

[0013] S4, robust optimization scheduling based on prediction results: based on the prediction results, establish a day-ahead scheduling robust optimization model considering the fluctuation interval, and calculate an optimal generation plan that is both safe and economic through model solving;

[0014] S5, rolling update: in the optimal generation plan put into the actual operation of the grid, in the intraday rolling optimization, use the latest data to repeat steps S3 and S4, adjust the generation plan, and continuously monitor the AI hybrid model and the day-ahead scheduling robust optimization model, if the performance decreases, trigger the model retraining mechanism;

[0015] S6, closed-loop feedback: update the end-to-end hybrid prediction model and the day-ahead scheduling robust optimization model to form a closed-loop system that continuously improves itself.

[0016] Preferably, in step S1, the load data is the total grid load data every hour in the past 3-5 years, the meteorological data is the hourly temperature, humidity, wind speed, solar intensity and weather forecast data during the same period, the date data is the holiday, weekday and weekend flag, and the grid data is the generator parameter and network topology structure, including minimum output, maximum output, ramp rate and cost coefficient.

[0017] Preferably, in step S2, the time feature is hour, day of the week, month and whether it is a holiday, the meteorological feature is current temperature, forecast temperature and apparent temperature, the historical load feature includes lag feature, rolling statistical feature and target variable, the lag feature is yesterday's same time load and last week's same time load, the rolling statistical feature is the average load of the past 3 hours and the maximum load of the past 24 hours, and the target variable is net load, net load = total load - new energy output.

[0018] Preferably, the specific steps of step S3 are as follows:

[0019] S31, use deep neural network to automatically learn the complex nonlinear relationship between static features and load, and output a vector representing the trend analysis result;

[0020] S32, use LSTM network to effectively capture complex time sequence patterns in load sequence by selectively remembering and forgetting information, and output a vector representing the summary of sequence situation;

[0021] S33, using the concatenation layer and the full connection layer, firstly, the trend vector output by step S31 and the sequence vector output by step S32 are simply spliced together in the concatenation layer to form a longer fusion feature vector, and then the spliced fusion feature vector is input into a smaller full connection network for further integration, so that the model learns how to balance the trend and sequence information, and finally an all-connection layer containing 3 neurons and without activation function is output, and the 3 neurons correspond to three outputs: pessimistic prediction value Q10, best point prediction value Q50 and optimistic prediction value Q90.

[0022] Preferably, the step S4 has the following specific steps:

[0023] S41, a day-ahead scheduling robust optimization model considering fluctuation interval is established:

[0024] (1) Optimization objective: ;

[0025] (2) Key constraint conditions:

[0026] (21) Power balance constraint: at any time, the total power generation must be equal to the predicted net load value;

[0027] (22) Unit operation constraint:

[0028] Upper and lower limits of power generation: the power generation of the unit must be within the minimum and maximum technical output range;

[0029] Ramp rate constraint: the speed of increasing or decreasing power output of the unit cannot exceed its maximum ramping capacity;

[0030] Minimum start-up and shutdown time constraint: once started, the unit must be continuously operated for at least a certain period of time before shutdown; once shut down, it must remain in the shutdown state for at least a certain period of time;

[0031] (23) Line transmission constraint: based on the direct current flow model, ensure that the power of any transmission line does not exceed its transmission limit;

[0032] (24) Reserve capacity constraint:

[0033] Upward reserve constraint: the system must reserve sufficient upward reserve capacity to cope with the maximum risk of upward fluctuation of the load;

[0034] Downward reserve constraint: the system must reserve sufficient downward reserve capacity to cope with the maximum risk of downward fluctuation of the load;

[0035] S42, model solving: using an efficient commercial solver or an open source solver for solving, the input being the point value Q50 of the future net load of the power grid within a certain time range and its fluctuation interval obtained in step S3 And the unit parameters, network parameters, the output is the start-stop state of each unit in the future same period, the planned output and the reserved standby capacity.

[0036] Due to the adoption of the above technical solutions, the application has the following advantages:

[0037] (1) The application adopts an end-to-end hybrid prediction model, and innovatively introduces quantile regression to output a prediction interval , which quantifies the uncertainty of the prediction, so that the scheduling optimization work no longer depends on a point prediction value, but can clearly know the range of possible load fluctuations, thereby changing from passive response to active defense, solving the scheduling risk problem caused by inaccurate prediction of the prior art;

[0038] (2) The application directly converts the prediction interval into a robust optimization constraint, which forces the system to reserve an upward reserve and a downward reserve to cover the interval, and within the known uncertainty range, the optimal cost scheduling scheme can be found, so that neither is it risky like traditional optimization, nor is it wasteful like extreme conservative scheduling, realizing optimal cost-effective safe scheduling, and solving the problem of mutual contradiction between economy and safety in the prior art;

[0039] (3) The application takes the net load as the object of prediction and optimization, directly incorporates the volatility of new energy into the uncertainty management framework, and reserves flexible standby capacity, so that the system has the ability to actively digest new energy fluctuations, thereby maximizing the acceptance of wind power and photovoltaic power under the premise of ensuring grid safety, promoting green transformation of new energy, and solving the problem of contradiction between clean energy consumption and grid stability in the prior art;

[0040] Overall, the application can realize the transition of grid scheduling from deterministic experience-based decision-making to probability-based and data-driven intelligent decision-making, and solve the economic and safety balance problem of grid scheduling in the face of uncertainty challenges under high proportion of renewable energy access, and has the advantages of high prediction accuracy, strong ability to cope with uncertainty, low total cost and strong practicality. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a step flowchart of the grid scheduling optimization method of the application.

[0042] Figure 2 is a step flowchart of the AI hybrid model prediction Q10 / Q50 / Q90Q of the application. DETAILED DESCRIPTION

[0043] The technical solutions of the application will be further specifically described below by examples in combination with the drawings.

[0044] As Figure 1 andFigure 2 As shown, the present application provides a power grid scheduling optimization method based on AI algorithm to predict power grid load changes, including the following steps:

[0045] S1, data preparation (preparation): collect the original data related to scheduling every hour in a certain time range, including load data, weather data, date data and power grid data, load data, weather data and date data are the basis for prediction, and power grid data is the basis for optimization.

[0046] Among them, the load data is the total load data of the power grid every hour in the past 3-5 years, through the analysis of these data, AI can learn the power utilization rule.

[0047] Among them, the weather data is the temperature, humidity, wind speed, solar intensity and weather forecast data every hour in the same period, and the weather is the most important external factor leading to load change; Among them, the temperature is closely related to the load, such as in summer, the higher the temperature, the more the air conditioner uses electricity, and the larger the load; The solar intensity will affect the photovoltaic power generation, thereby changing the net load of the power grid; Therefore, the AI model needs to learn the relationship that "how much will the load increase when the temperature rises by one degree".

[0048] Among them, the date data is the holiday, weekday and weekend mark, and people's power utilization habit is also closely related to the date type, generally the power utilization curve of weekday is completely different from that of weekend, and the power utilization mode of holiday is more special, therefore, the date label can help the AI model understand the social activity background of a day, so as to make more accurate judgment.

[0049] Among them, the power grid data is the generator set parameter (minimum output, maximum output, climbing rate and cost coefficient) and network topology structure (physical connection diagram of power grid, including transformer substation, transmission line and its capacity limit), after the AI predicts the load, the optimization algorithm needs to formulate the power generation plan according to these hardware limits, and select the power generation combination with the lowest total cost.

[0050] S2, feature engineering (processing): through extraction, calculation and combination, the collected scheduling related original data is converted into key signals and indexes that can be understood by AI model, including time features, weather features and historical load features.

[0051] Among them, the time feature is hour, day of the week, month and whether it is a holiday, which is used to directly provide a time stamp for AI.

[0052] Among them, the weather feature is the current temperature, forecast temperature and apparent temperature, the apparent temperature is a derived feature, which is generally calculated by temperature and humidity, such as the apparent temperature can reach 35℃ when the temperature is 30℃ due to high humidity, the combination of the two can reflect the hot feeling of human body better than 30℃ alone, so the prediction of load is more accurate.

[0053] Among them, the historical load characteristics include lag features, rolling statistical features and target variables. The lag features are the load at the same time yesterday and the load at the same time last week, which are used for AI to remember key past moments; the rolling statistical features are the average load in the past 3 hours and the maximum load in the past 24 hours, which are used for AI to summarize the recent situation; and the target variable is the net load, net load = total load - new energy output, which is used to redefine the prediction target, so as to make the optimization more accurate and efficient, and greatly promote the consumption of new energy.

[0054] S3, AI model prediction (prediction) based on quantile regression: build an end-to-end hybrid prediction model to accurately predict the point value Q50 of the future grid net load and its fluctuation interval at one time The specific steps are as follows:

[0055] S31, use deep neural network to automatically learn the complex nonlinear relationship between static features and load, and output a vector representing the trend analysis result;

[0056] Input layer: the number of neurons is equal to the number of static features, such as hours, days, months, etc. There are 10 neurons in total.

[0057] Hidden layer: 1 to 2 layers can be used; the first hidden layer contains 64 or 128 neurons, and uses ReLU activation function to introduce nonlinearity; the second hidden layer contains 32 or 64 neurons, and also uses ReLU activation function.

[0058] Output layer: the output of this branch is a vector representing the trend analysis result, such as a 32-dimensional vector.

[0059] S32, use LSTM network to effectively capture complex time sequence patterns in load sequence by selectively remembering and forgetting information, and output a vector representing sequence situation summary;

[0060] Input layer: receives a matrix with a dimension of 72 represents the time step (past 72 hours), and 2 represents the feature vector at each time point (historical net load / historical temperature);

[0061] LSTM layer: 1 to 2 layers can be stacked; the first layer LSTM contains 64 or 128 memory cells and returns the output of the entire time sequence to connect the next layer LSTM; the second layer LSTM contains 32 or 64 memory cells and returns the output of the last time step, i.e. outputs a vector containing long-term dependence information of the entire sequence, such as a 32-dimensional vector;

[0062] Output layer: the output of this branch is a vector representing the trend analysis result.

[0063] S33, using the concatenation layer and the full connection layer, firstly, the trend vector (such as 32 dimensions) output by step S31 and the sequence vector (such as 32 dimensions) output by step S32 are simply concatenated together in the concatenation layer to form a longer fusion feature vector (such as 64 dimensions), then the concatenated fusion feature vector is input into a smaller full connection network (such as containing 16 neurons) for further integration, so that the model learns how to weigh the trend and sequence information, and finally an output of a full connection layer containing 3 neurons and without activation function is output, and the 3 neurons correspond to three outputs: the pessimistic prediction value Q10, the best point prediction value Q50 and the optimistic prediction value Q90.

[0064] S4, robust optimization scheduling (decision) based on the prediction result: based on the prediction result, a day-ahead scheduling robust optimization model considering the fluctuation interval is established, and an optimal generation plan which is both safe and economic is calculated through model solving, and the specific steps are as follows:

[0065] S41, establishing a day-ahead scheduling robust optimization model considering the fluctuation interval:

[0066] (1) optimization objective: , The calculation formula is:

[0067] ;

[0068] In the formula, T is the total time period of the scheduling period, such as 24 hours; N is the number of generating units; is the planned generation power of unit i at time period t; is the fuel cost when the generation power of unit i is ; is a binary variable, =1 indicates that unit i starts at time period t; is the start-up cost of unit i; is a binary variable, =1 indicates that unit i stops at time period t; is the shutdown cost of unit i; is the up-regulation reserve capacity reserved for unit i at time period t; is the price (unit capacity cost) of the up-regulation reserve; is the down-regulation reserve capacity reserved for unit i at time period t; is the price (unit capacity cost) of the down-regulation reserve;

[0069] (2) key constraint conditions:

[0070] (21) power balance constraint: at any time, the total generation power must be equal to the predicted net load value, and the calculation formula is:

[0071] ;

[0072] where, is the net load point forecast value at time period t, i.e., the best point forecast value Q50;

[0073] (22) Unit operation constraints:

[0074] (a) Generation output upper and lower limit constraints: The generation output of a unit must be within the minimum and maximum technical output range, calculated as:

[0075] ;

[0076] (b) Ramp rate constraints: The rate of increase or decrease of the unit output cannot exceed its maximum ramping capability, calculated as:

[0077] Upward ramping constraints: ;

[0078] Downward ramping constraints: ;

[0079] (c) Minimum start-up and shut-down time constraints: Once a unit is started, it must be continuously operated for at least a certain period of time before it can be shut down; once it is shut down, it must remain in the off state for at least a certain period of time, calculated as:

[0080] Minimum operating time constraints: ;

[0081] Minimum shut-down time constraints: ;

[0082] where, and are the minimum and maximum technical outputs of unit i; and are the upward and downward ramp rates of unit i; and are the minimum operating time and minimum shut-down time of unit i;

[0083] (23) Line transmission constraints: Based on the DC power flow model, ensure that the power of any transmission line does not exceed its transmission limit, calculated as:

[0084] ;

[0085] where, is the maximum transmission capacity of line l; is the generation shift distribution factor, representing the influence coefficient of the output change of unit i on the power of line l;

[0086] (24) Reserve capacity constraints:

[0087] (a) Upward reserve constraint: the system must reserve sufficient upward reserve capacity to cope with the maximum risk of load fluctuation upward, the calculation formula is:

[0088] System-level constraints: ;

[0089] Unit-level constraints:

[0090] (b) Downward reserve constraint: the system must reserve sufficient downward reserve capacity to cope with the maximum risk of load fluctuation downward.

[0091] System-level constraints: ;

[0092] Unit-level constraints: ;

[0093] In the formula, is the 90th percentile of the net load in period t (i.e. the optimistic prediction value Q90); is the 10th percentile of the net load in period t (i.e. the pessimistic prediction value Q10).

[0094] S42, model solving: use efficient commercial solvers (such as Guiobi, CPLEX) or open source solvers (such as CBC) for solving, input is the point value Q50 of the future 24 hours of grid net load and its fluctuation interval obtained in step S3, as well as unit parameters, network parameters, output is the start-stop state, planned output and reserved reserve capacity of each unit in the future 24 hours.

[0095] S5, rolling update (execution): the above day-ahead scheduling is executed once a day in the optimal generation plan put into the actual operation of the grid, but as time goes by, there will be more accurate short-term and ultra-short-term load forecasts, so rolling optimization will be carried out within the day, using the latest data to repeat steps S3 and S4 to adjust the generation plan; at the same time, the error of the AI hybrid model and the actual cost of the optimization result of the day-ahead scheduling robust optimization model are continuously monitored, and if the performance decreases, the model retraining mechanism is triggered.

[0096] S6, closed-loop feedback (growth): the actual occurring load data and unit operation state data are fed back to the database for updating the end-to-end hybrid prediction model and the day-ahead scheduling robust optimization model, forming a closed-loop system for continuous self-improvement.

[0097] In summary, the present application uses an AI model to make high-precision predictions of future grid net load, and uses the prediction results and their uncertainties as key inputs to drive an optimization model that considers safety constraints to generate an economic, reliable and low-carbon generation plan for use in a data-driven intelligent grid dispatching optimization method.

[0098] Finally, it should be noted that the above examples are merely intended to illustrate the technical solutions of the present application, and are not intended to limit the same. Although the present application has been described in detail with reference to the above examples, those of ordinary skill in the art should understand that the specific embodiments of the present application can still be modified or equivalently replaced without departing from the spirit and scope of the present application, and any modification or equivalent replacement that does not depart from the spirit and scope of the present application should be covered in the scope of the claims of the present application.

Claims

1. A power grid scheduling optimization method based on AI algorithm for predicting power grid load changes, characterized in that, Comprise the following steps: S1, data preparation: collect the hourly scheduling related raw data in a certain time range in the past, including load data, weather data, date data and power grid data; S2, feature engineering: through extraction, calculation and combination, the collected scheduling related raw data is converted into key signals and indicators that can be understood by AI models, including time features, weather features and historical load features; S3, AI model prediction based on quantile regression: build an end-to-end hybrid prediction model to accurately predict the point value Q50 of the future power grid net load and its fluctuation interval at one time ; S4, robust optimization scheduling based on prediction results: based on the prediction results, a day-ahead scheduling robust optimization model considering the fluctuation interval is established, and a safe and economic optimal generation plan is calculated through model solving; S5, rolling update: in the actual operation of the optimal generation plan in the power grid, the latest data is used to repeat steps S3 and S4 to adjust the generation plan, and the AI hybrid model and the day-ahead scheduling robust optimization model are continuously monitored, and if the performance decreases, the model retraining mechanism is triggered; S6, closed-loop feedback: update the end-to-end hybrid prediction model and the day-ahead scheduling robust optimization model to form a closed-loop system for continuous self-improvement. 2.The power grid scheduling optimization method based on AI algorithm for predicting power grid load change according to claim 1, characterized in that: In step S1, the load data is the hourly total load data of the power grid in the past 3-5 years, the weather data is the hourly temperature, humidity, wind speed, solar intensity and weather forecast data in the same period, the date data is the holiday, weekday and weekend marker, and the power grid data is the generator set parameter and network topology structure, including minimum output, maximum output, ramp rate and cost coefficient. 3.The power grid scheduling optimization method based on AI algorithm for predicting power grid load change according to claim 1, characterized in that: In step S2, the time feature is hour, day of the week, month and whether it is a holiday, the weather feature is current temperature, forecast temperature and apparent temperature, the historical load feature includes lag feature, rolling statistical feature and target variable, the lag feature is yesterday's same time load and last week's same time load, the rolling statistical feature is past 3 hours average load and past 24 hours maximum load, and the target variable is net load, net load = total load - new energy output. 4.The power grid scheduling optimization method based on AI algorithm for predicting power grid load change according to claim 1, characterized in that, The specific steps of step S3 are as follows: S31, a deep neural network is used to automatically learn the complex nonlinear relationship between static features and load, and output a vector representing the trend analysis result; S32, an LSTM network is used to effectively capture complex time sequence patterns in the load sequence by selectively remembering and forgetting information, and output a vector representing the summary of sequence trend; S33, a concatenation layer and a fully connected layer are used, first, the trend vector output by step S31 and the sequence vector output by step S32 are simply concatenated in the concatenation layer to form a longer fusion feature vector, then the concatenated fusion feature vector is input into a smaller fully connected network for further integration, and the model learns how to balance trend and sequence information, and finally a fully connected layer with 3 neurons and without activation function is output, and the 3 neurons correspond to three outputs: pessimistic prediction value Q10, best point prediction value Q50 and optimistic prediction value Q90. 5.The power grid scheduling optimization method based on AI algorithm for predicting power grid load change according to claim 1, characterized in that, The specific steps of step S4 are as follows: S41, a day-ahead scheduling robust optimization model considering the fluctuation interval is established: (1) Optimization objectives: ; (2) Key constraint conditions: (21) Power balance constraint: At any time, the total generation power must equal the predicted net load value; (22) Unit operation constraints: Generation output upper and lower limit constraint: The generation power of a unit must be within the minimum and maximum technical output range; Ramp rate constraint: The speed of increasing or decreasing the output of a unit cannot exceed its maximum ramping capability; Minimum up and down time constraint: Once a unit is started, it must be continuously operated for at least a certain period before it can be shut down; once it is shut down, it must remain in the shutdown state for at least a certain period; (23) Line transmission constraint: Based on the DC power flow model, ensure that the power of any transmission line does not exceed its transmission limit; (24) Reserve capacity constraints: Up-regulation reserve constraint: The system must reserve sufficient up-regulation reserve capacity to address the maximum risk of upward fluctuation of the load; Down-regulation reserve constraint: The system must reserve sufficient down-regulation reserve capacity to address the maximum risk of downward fluctuation of the load; S42, model solving: using efficient commercial solver or open source solver for solving, input is the point value Q50 of the future certain time range of the power grid net load and its fluctuation interval obtained in step S3 and unit parameters, network parameters, and output is the start-stop state, planned output and reserved standby capacity of each unit in the future same period.

Images