A Day-Ahead Dispatching Method for Power Systems Based on Conv-Seq2Seq Model in Flexible Environment

Abstract

The present invention belongs to the technical field of power system dispatch optimization, and specifically relates to a method for day-ahead dispatch of a power system based on a Conv‑Seq2Seq model in a flexible environment. The method comprises the following steps: Step 1: establishing a model for traditional thermal power units, deep peak-shaving units and load-side elastic resources in the power system, wherein the load-side elastic resources include curtailable loads and shiftable loads; Step 2: constructing a data set based on load forecast data, wind power forecast data and other information and a corresponding dispatch plan; Step 3: constructing a deep learning model based on Conv‑Seq2Seq; Step 4: using the data set to train the deep learning model to obtain a trained deep learning decision model, and inputting the load forecast information and wind power forecast information to be decided into the deep learning model for decision-making, and outputting a dispatch plan; Step 5: performing auxiliary decision-making correction on the output scheme of the deep learning model to obtain a safe and feasible solution that meets the constraints.

Claims

1. A method for day-ahead dispatch of a power system based on a Conv-Seq2Seq model in a flexible environment, characterized by: The method comprises the following steps: Step 1: Build a model for traditional thermal power units, deep peak-shaving units, and load-side elastic resources in the power system, where the load-side elastic resources include curtailable loads and shiftable loads; The elastic resources of the power system on the power supply side include N fuel Thermal power units and N w Typhoon power generation units, among which thermal power generation units are divided into traditional thermal power generation units and deep peaking generation units according to their peaking capacity. Deep peaking generation units have N dpr The elastic resources on the load side include rigid load and flexible load, among which rigid load is the load that must meet the user's electricity demand and does not participate in scheduling; flexible load includes shiftable load and curtailable load; Step 2: Construct a data set based on load forecast data, wind power forecast data and other information as well as the corresponding dispatch plan; Step 2.1: Establish the objective function: Where: N fuel Indicates that the thermal power unit has N fuel Units, of which deep peak-shaving units are N dpr tower, is the total operating cost of the i-th traditional thermal power unit in period t, is the total operating cost of the g-th deep peak-shaving unit in period t, C cut and C sh represent the user compensation costs for loads that can be curtailed and loads that can be shifted, respectively; Step 3: Build a deep learning model based on Conv-Seq2Seq; Step 4: Use the data set to train the deep learning model to obtain a trained deep learning decision model. Input the load forecast information and wind power forecast information to be decided into the deep learning model for decision-making and output the scheduling plan. Step 5: Perform auxiliary decision-making corrections on the output scheme of the deep learning model to obtain a safe and feasible solution that meets the constraints.

2. The method for day-ahead dispatch of a power system based on a Conv-Seq2Seq model in a flexible environment according to claim 1, characterized in that: The scheduling plan includes the start and stop status and power generation capacity of traditional thermal power units and deep peak-shaving units, the start time and compensation price of load that can be shifted, and the load reduction amount and compensation price of load that can be reduced.

3. The method for day-ahead dispatch of a power system based on a Conv-Seq2Seq model in a flexible environment according to claim 1, characterized in that: The step 1 specifically includes: Step 1.1: Build a cost model for traditional thermal power plants: The operating costs of traditional thermal power units include coal consumption costs and start-stop costs They are Where: a i ,b i ,c i is the coal consumption coefficient of the i-th thermal power unit; P i,t is the output of the i-th traditional thermal power unit in period t; ΔT is the time interval; Indicates whether the thermal power unit is started. When the genset is in working state; and are the startup cost and shutdown cost of the i-th thermal power unit respectively; Therefore, the total cost of traditional thermal power units for Step 1.2: Build a cost model for deep peaking units: When the deep peaking unit operates in the conventional peaking state, its cost model is consistent with the cost model of the traditional thermal power unit; when the deep peaking unit operates in the deep peaking state, it can avoid frequent unit startup, but it will increase the additional coal consumption cost caused by the unit output deviating from the rated value. and the life loss cost caused by irreversible damage due to unit output deviation from rated value Specifically, it can be expressed as: Where: Indicates the deep peak regulation state of thermal power units. Indicates that the unit is in deep peak regulation state. Indicates that the unit is operating in a non-deep peak-shaving state; and v g are the coal consumption rate coefficients of the g-th thermal power unit in the deep peak regulation state and the conventional minimum technical output state respectively; ε g is the coal consumption rate of the g-th thermal power unit at rated output; P g,t is the output of the g-th deep peak-shaving unit during period t; ρ coal is the unit coal price; N g,t is the number of cycles of rotor cracking of the g-th thermal power unit, and its value is related to P g,t is closely related; ω is the operating loss coefficient of the deep peak-shaving unit; is the purchase cost of the g-th unit; Therefore, the total cost of deep peak load regulation units for Step 1.3: Build curtailable load model: Reducible load refers to the load that can be partially reduced in operating power while meeting the basic needs of users. The value range of can be expressed as Where: is the rated power of the load that can be reduced; is the load elasticity coefficient that can be reduced during period t, which is related to the compensation price. Represents the upper limit of the load reduction that can be taken into account under the current compensation price; User compensation cost after scheduling C cut for Where: is the compensation price for unit power load reduction during period t; Reducible load reduction factor Reflecting the elasticity of load reduction, the price of load reduction compensation will affect the amount of load reduction. When the price is at a low level, users have no response; when the price increases to a certain value, When the price increases, the users begin to accept the compensation and are willing to reduce the electricity load. As the price increases, the amount of load that can be reduced also increases gradually. When the price increases to a certain value, When the elastic resources are fully exploited, the load reduction amount will no longer increase, and the load reduction coefficient It can be expressed as Where: cut Indicates the user's sensitivity to the compensation of the curtailable load. The smaller it is, the more sensitive the user is to it. Refers to the maximum spring rate that can reduce the load, and Step 1.4: Create a translatable load model: A shiftable load refers to a load that has a fixed working time and can be shifted as a whole during the working period. Appropriate time should be selected for shifting to achieve the effect of peak shaving and valley filling. The acceptable shiftable range of a shiftable load is is the compensation price for unit power load shift, and the continuous operation time constraint that needs to be met is Where: t sh is the starting period after the load shifts; T sh is the duration of the load that can be translated; is a 0-1 state variable indicating whether a translation occurs during the t period, Indicates that the load has shifted to period t; The load power that can be shifted during period t for Where: L sh is the rated power of the load that can be translated; Economic compensation for users C sh for Where: is the compensation price for unit power load shift; When the price is at a low level, users do not respond; when the price increases to a certain value When the price increases, the acceptable shift period also increases gradually. When the price increases to a certain value, the user begins to accept compensation and the acceptable shift period begins to increase. When , the elastic resources are fully exploited and the acceptable shift period does not change; Therefore, its model can be expressed as Where: is the original operation starting period of the movable load, The maximum number of extended periods during which the shiftable load can accept the shift period, ρ sh is the elastic expansion coefficient of the acceptable translation period; ξ sh is the compensation sensitivity of the translatable load.

4. The method for day-ahead dispatch of a power system based on a Conv-Seq2Seq model in a flexible environment according to claim 1, characterized in that: The step 2 specifically includes the following steps: Step 2.2: Create constraints: The constraints mainly include power balance constraints, upper and lower output limits of thermal power units, thermal power unit ramp constraints, and line transmission capacity constraints, as shown in the following formula: The power balance constraint of the entire system during period t is Where: P i,t and are the outputs of the i-th thermal power unit and the j-th wind power unit in period t, respectively, P t load Indicates the total load value during period t, P t cut is the load reduction power that can be reduced during period t, P t sh* and P t sh are the power consumption of the load that can be shifted before and after the scheduling in period t, P t rigid is the rigid load power during period t, In order to reduce the load power before participating in the dispatch, The system line transmission capacity constraint during period t is: Where: and is the power transmission distribution coefficient of each node in the system to line br, and the superscripts fuel, w, and load represent the nodes where thermal power, wind power, and load are located; N b is the number of grid nodes; is the load forecast value at node g in the system during period t after flexible load scheduling, and is the upper limit of the power flow of line br, The upper and lower limits of thermal power unit output are Where: P i min ,P i max They are the upper and lower limits of the generator set output range respectively. For conventional units, P i min is the conventional minimum technical output, while for deep peak load units, P i min It is the minimum output in the deep peak regulation stage; Indicates whether the thermal power unit is started. When the genset is in working state, The climbing constraint of thermal power unit is Where: r i down ,r i up are the downward and upward climbing rates of the i-th thermal power unit, ΔT is the time interval, The start and stop time constraints of thermal power units are Where: Respectively represent the number of continuous operation periods and continuous shutdown periods of the i-th thermal power unit, T i on ,T i off is the minimum number of operating hours and minimum number of downtime hours that the i-th thermal power unit must meet; Step 2.3: Based on the load forecast data, wind power forecast data and other information, with the operating equipment and load in the power system as constraints and the goal of minimizing the system operating cost, use the solver to generate a scheduling plan corresponding to the load forecast data and wind power forecast data. Finally, a data set is constructed based on the load forecast data, wind power forecast data and other information and the corresponding scheduling plan.

5. The method for day-ahead dispatch of a power system based on a Conv-Seq2Seq model in a flexible environment according to claim 1, characterized in that: Step 3 constructs a deep learning model based on Conv-Seq2Seq, which specifically includes the following steps: Step 3.1: Construct the deep learning model input sequence, preprocess the dataset generated in step 2, and obtain the input sequence X = (x1, x2, ..., x T ), since the multi-layer convolutional network used in the encoder structure cannot reflect the position information in the sequence, the position encoding L=(l1,l2,…,l T ) to represent the absolute position of the sequence element in the sequence, and the comprehensive input E=(x1+l1,x2+l2,…,x T +l T ); Step 3.2: Create the encoder structure, through n block The layer convolution block extracts information from the input. Since the multi-layer convolution blocks have the same structure, only the single-layer convolution block is introduced. The single-layer convolution block mainly includes a convolutional network, a gated linear unit (GLU) activation function, and a residual connection. The details are as follows: For each convolution kernel of size k, its input is k elements embedded in d-dimensional space Its formula is Y conv =F(X conv )+b conv Where: Parameters of the convolutional network The output dimension is twice that of the input, Using GLU as the nonlinear activation function, its formula is Where: Φ represents the GLU function, represents the element-by-element multiplication of matrices, Its dimension is Y conv Half of ; σ(·) is the gate unit related to A; the residual connection is introduced to obtain the output of a convolutional network, and its formula is Where: n represents the nth layer network, n≤n block ; is the t-th input of the n-th layer convolutional network; Between different layers of convolution blocks, the input of the mth layer convolution block is the output of the m-1th layer convolution block, where 1<m≤n block , and process the input sequence through the Padding operation to ensure the consistency of the input and output sequence lengths of different layers of convolution blocks; the nth block The output of the convolutional block is processed by a fully connected layer, and the result is used as the initial hidden state of the decoder. The value of Step 3.3: Introduce the Attention mechanism in the decoding part, so that the decoder focuses on different information at different times, avoiding information compression and improving information utilization; Step 3.3.1: Calculate the influence e of the k-th input sequence element of the encoder on the t-th input sequence element of the decoder k,t : Where, F en Represents the encoder encoding operation, V, W and U are the parameters that can be trained by the neural network, is the hidden state of the encoder at step t, is the hidden state of the decoder at step t; Step 3.3.2: Set the impact level e k,t After softmax processing, the weight a of each hidden state is obtained k,t : Step 3.3.3: According to the weights a of each hidden state k,t And linearly add the product of each hidden state and its corresponding weight to get the semantic vector V t : While avoiding information compression, the model pays different attention to each moment during decoding, improving the ability to utilize information; Step 3.4: Input the information extracted by the encoder into the decoder for decoding and output the schedule. The decoder network is composed of a GRU network. The decoder receives the last hidden state record of the encoder. as its initial hidden state and generates the output sequence Therefore, the output of the decoder at step t+1 It can be expressed as Where: F de Represents the decoder decoding operation, is the hidden state of the decoder at step t, Decoder hidden state It can be expressed as: Where: F de,h Represents the decoder to calculate the hidden state operation; In the process of generating the output sequence, the network output P at step t t out As the network input for step t+1, in this operation, the network outputs P t out There is a prediction error. As time steps move, the error accumulates. In order to make the network perform better in dealing with the prediction error, consider using the network output P of step t with a certain probability. t out The label value is used as the network input for step t+1, which can be expressed as Where: is the input of the decoder at step t+1; P t out is the output generated by the decoder at step t, The network output P for step t t out label value; ε is a random number generated to satisfy a uniform distribution between 0 and 1. When it is less than 0.5, the input of the decoder at step t+1 is the corresponding label value; when it is greater than or equal to 0.5, the input of the decoder at step t+1 is the output generated by the decoder at step t.

6. The method for day-ahead dispatch of a power system based on a Conv-Seq2Seq model in a flexible environment according to claim 1, characterized in that: Step 4 of deep learning model training and obtaining a scheduling plan through the trained model specifically includes the following steps: Step 4.1: Initialize the deep learning model network parameters, where epoch is the number of current cycles, EPOCH is the total number of cycles, and val_loss is the validation set loss function; Step 4.2: Batch input sample data, calculate the output sequence through the Conv-Seq2Seq model, and generate a scheduling plan; Step 4.3: Calculate the loss function according to the output scheduling plan in combination with the tag values, and update the network parameters through the ADAM algorithm; and evaluate the training status of the deep learning model through the validation set, and save the parameters of the deep learning model with lower val_loss as the optimal parameters; Step 4.4: epoch = epoch + 1. If epoch < EPOCH, jump to Step 4.2; Step 4.5: Generate a scheduling plan corresponding to the load prediction data to be decided and the wind speed prediction data under the optimal parameters.

7. The method for day-ahead dispatch of a power system based on a Conv-Seq2Seq model in a flexible environment according to claim 1, characterized in that: The auxiliary decision-making correction in Step 5 specifically includes the following steps: Use the output scheme of the deep learning model as the input of the auxiliary decision-making correction, fix the start-stop scheme of the units therein, and make corrections with the goal of minimizing the system operation cost, and finally obtain a safe and economic scheduling plan.

Images