Water, wind and light power multi-target interval forecasting method considering forecasting error complementarity

Abstract

The invention provides a water-wind-light power multi-target interval forecasting method considering forecasting error complementarity, and the method comprises the steps: arranging the historical monthly scale data of a selected water-wind-light complementary system, including water-electricity power time sequence data, wind power time sequence data, photoelectric power time sequence data, total power time sequence data and various climate system indexes; the method comprises the following steps of: calculating correlation among a climate system index, hydroelectric power, wind power, photoelectric power and total power by adopting a maximum mutual information coefficient method in consideration of various time lags, screening forecast factors corresponding to the power on the basis, and screening the forecast factors corresponding to the power in consideration of complementarity among the hydroelectric power, the wind power and the photoelectric power. And constructing a point forecasting and interval forecasting model based on a deep learning model, performing accumulated forecasting on the total power after performing independent forecasting on each power, and performing combined forecasting on the total power after accumulating each power. According to the method, the complementarity of forecasting errors is considered, so that the forecasting precision of each power of the water-wind-light system is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power forecasting for a water-wind-solar complementary system, and specifically relates to a multi-objective interval forecasting method for water-wind-solar power that takes into account the complementarity of forecast errors. Background Art

[0002] my country's energy structure is constantly transforming towards green and sustainable development to achieve the "dual carbon" goal. Therefore, carbon emissions can be reduced by building a new power system based on renewable energy such as wind and light. However, the significant randomness and volatility of wind power and photovoltaic power make it difficult to absorb new energy, and the problems of "abandoning wind" and "abandoning light" are very prominent. Implementing the complementary operation of water, wind and light is an important way to solve this problem.

[0003] Hydropower is the renewable energy technology with the largest development scale and the most mature engineering technology in my country. Meanwhile, wind power and photovoltaic power generation will be the main incremental installed capacity of my country's power system in the future. Therefore, how to safely and efficiently utilize the water, wind and light resources in the river basin, give full play to their important role in carbon emission reduction, and promote sustainable economic and social development is a consensus problem that the global energy field is seeking to break through.

[0004] The traditional method predicts hydropower, wind power and photovoltaic power separately, and then accumulates them to get the total power of the hydropower, wind power and photovoltaic power system. This method has the problem of easy error accumulation and does not consider the temporal and spatial complementarity of hydropower, wind power and photovoltaic power. In order to improve the accuracy of system power forecast, the temporal and spatial correlation and complementarity are first considered, and the forecast factors are selected from the climate system index and power. The maximum NSE and the maximum CRFE are used as the objective function, and the forecast is made on this basis to achieve the best forecast effect. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-objective medium- and long-term interval forecasting method for each power of water, wind and solar power taking into account the complementarity of forecast errors, so as to solve the problem proposed in the above background technology that a multi-objective medium- and long-term interval forecasting method for each power of water, wind and solar power taking into account the complementarity of forecast errors is proposed, so as to obtain a more accurate forecasting effect, thereby providing an important basis for the operation and management of the power system.

[0006] To achieve the above object, the present invention provides the following technical solution: a multi-objective interval forecasting method for water, wind and solar power considering the complementarity of forecast errors, the method comprising the following steps:

[0007] Step 1: Organize the historical monthly data of the selected water-wind-solar complementary system, including hydropower time series data, wind power time series data, photovoltaic power time series data, total power time series data and various climate system indexes;

[0008] Step 2: Considering the time lag effect of the prediction factor on the predicted power, the maximum mutual information coefficient method is used to calculate the correlation between each climate system index and each power, and based on this, the prediction factor corresponding to each power is selected from a large number of climate system indices;

[0009] Step 3: Analyze the complementarity between hydropower, wind power and photovoltaic power in the water-wind-solar multi-energy complementary system;

[0010] Step 4: construct various power forecasting models based on the long short-term memory network model, construct an optimization model with the maximum Nash efficiency coefficient and the maximum prediction error complementary rate CRFE as the objective function, use the sparrow optimization algorithm SSA to solve, optimize the parameters (learning rate, regularization coefficient, number of hidden nodes) in the LSTM model, and make a point forecast of the system power;

[0011] Step 5: Use the Bootstrap resampling method to calculate the residuals between the point prediction values ​​and the actual values ​​generated by the LSTM model for each power, resample the residuals to generate multiple sets of new residual data, and superimpose these resampled residuals on the point prediction values ​​to generate multiple sets of new prediction results. Then, calculate the upper and lower limits of the prediction interval based on these new prediction results to achieve interval prediction. The evaluation indicators of the model use interval coverage and interval standardized average width, and the comprehensive criterion of coverage width as evaluation indicators.

[0012] Preferably, in step 2, the maximum mutual information coefficient is an indicator for measuring the correlation between two variables based on mutual information, which can measure the amount of information contained in one random variable about another random variable. The calculation method is:

[0013] The scatter plot consisting of the power forecast object np and the climate system index (or power) X in the data set is gridded, and the mutual information value of each grid is calculated:

[0014]

[0015] In the formula, p(x) and p(n p) is the marginal probability density function of the two variables; p(x,n p ) is the probability density function of the two variables,

[0016] The maximum mutual information value I*(X,N p )Normalization:

[0017]

[0018] Where: a and b are the number of rows and columns of the grid respectively; M(D) a,b is the normalized maximum mutual information value under fixed grid partitioning.

[0019] Perform grid division for different scenarios and repeat the above steps to obtain all M(D). a,b The maximum value is the maximum mutual information coefficient MIC:

[0020]

[0021] In the formula: ab < B(T) is the constraint condition; B(T) is usually taken as the 0.6th power of the dataset.

[0022] Preferably, in step 3, quantify the complementarity of hydropower, wind power, and photovoltaic power. Adopt the MIC method to calculate the mutual information index to quantify the complementarity. According to the analysis results, evaluate the complementarity between different energy sources, and use mutual information to measure the non-linear complementarity.

[0023] Preferably, in step 4, construct a joint prediction model and an accumulative prediction model for each power based on the LSTM model. For the joint prediction model, use the climate system index corresponding to the power to be predicted as the input and the power to be predicted as the output; for the accumulative prediction model, use the climate system index corresponding to each power as the input and the power to be predicted as the output respectively. After establishing three prediction models, accumulate the prediction results to obtain the output result of the accumulative prediction model of the total power.

[0024] The principle of the LSTM model is as follows: LSTM (Long Short-Term Memory) is a special recurrent neural network (RNN). It solves the problems of gradient vanishing and gradient explosion that occur in traditional RNNs when dealing with long time series by introducing a gating mechanism. The basic unit of the LSTM network includes three key gates: the input gate, the forget gate, and the output gate. These gate mechanisms make the LSTM perform particularly well in capturing long-term dependencies in time series. The LSTM model formula is as follows:

[0025] Input gate:

[0026] i t = σ(W i ·[h t-1 , X t +b i )

[0027] In the formula: i t is the input gate vector, which determines which new information will be stored in the cell state; σ is the Sigmoid activation function, which limits the output value between 0 and 1; W i is the weight matrix of the input gate; h t-1 is the hidden state of the previous time step; X tis the input vector of the current time step (including climate system indices of multiple time steps); b i Bias vector for the input gate.

[0028] Forget Gate:

[0029] f t =σ(W f ·[h t-1 ,X t ]+b f )

[0030] Where: f t is the forget gate vector, which determines which information is forgotten from the cell state; W f is the weight matrix of the forget gate; b f is the bias vector of the forget gate.

[0031] Candidate memory cell status:

[0032]

[0033] Where: is the candidate memory cell state, indicating a new candidate memory; tanh is the hyperbolic tangent activation function, which limits the output value to between -1 and 1; W C is the weight matrix of the candidate memory cell state; b C is the bias vector of the candidate memory cell state.

[0034] Memory cell status:

[0035]

[0036] Where: C t is the memory cell state at the current time step; f t is the forget gate vector; * is element-by-element multiplication, which does not need to meet the dimension requirements of matrix multiplication during calculation; C t-1 is the memory cell state at the previous time step.

[0037] Output Gate:

[0038] o t =σ(W o ·[h t-1 ,X t ]+b o )

[0039] Where: t is the output gate vector, which determines which parts of the cell state are output; W o is the weight matrix of the output gate; b o is the bias vector of the output gate.

[0040] Hidden state:

[0041] h t =o t *tanh(C t )

[0042] Where: h t is the hidden state at the current time step.

[0043] Construct an LSTM key parameter optimization model (learning rate, regularization coefficient, number of hidden nodes), with the maximum Nash-Sutcliffe efficiency coefficient (NSE) and the maximum complementary rate of forecast error CRFE (Complementary Rate of Forecast Error) as the objective function, and the parameter value range as the constraint condition. The Nash efficiency coefficient calculation formula is as follows:

[0044]

[0045] Where: is the point prediction value of the corresponding power in the tth period; N t is the actual value of the corresponding power in the tth period; T is the total number of time periods of the sample data; is the mean value of power.

[0046] The calculation formula of forecast error complementarity rate is as follows:

[0047]

[0048] Where: M is the total number of samples; M p is the number of hydropower, wind and solar power stations; is the jth predicted value of the ith sample; N i,j is the jth actual value of the ith sample. The forecast error complementarity rate is obtained by normalizing the difference between the numerator and the denominator to a value between 0 and 1. The closer CRFE is to 1, the smaller the model prediction error is, that is, the better the complementarity between different powers is; the closer it is to 0, the larger the model prediction error is, that is, the poorer the complementarity between different powers is.

[0049] The sparrow optimization algorithm SSA multi-objective function is used to solve the problem, optimize the parameters in LSTM (learning rate, regularization coefficient, number of hidden nodes), and predict the system power. The sparrow optimization algorithm (SSA) is used to solve the single-objective optimization problem, and the weighted summation method is used to merge the two objectives into one objective function. The specific steps are as follows:

[0050] Set the population size N and the maximum number of iterations T. Initialize the population position X:

[0051]

[0052] Where: X ij is the position of the i-th sparrow in the j-th dimension; and is the lower and upper bounds of the jth dimension; rand() generates a random number between 0 and 1; d is the time step;

[0053] Calculate the fitness of each sparrow:

[0054] f(X i )=w1·NSE(X i )+w2·CRFE(X i )

[0055] Where: w1 and w2 are the weight coefficients of NSE and CRFE; NSE(X i ) is the NSE value corresponding to the i-th sparrow position; CRFE(X i ) is the CRFE value corresponding to the i-th sparrow position.

[0056] Sort by fitness value and select the top N L The sparrow acts as the leader and the rest as followers. Leader position update:

[0057]

[0058] Where: is the position of the i-th leader at the t+1th iteration; is the position of the i-th leader at the t-th iteration; λ is a random number.

[0059] Follower Location Updates:

[0060]

[0061] is the position of the i-th follower at the t+1th iteration; is the position of the ith follower at the tth iteration; β is a random number; The best position for the navigator.

[0062] By comparing the fitness of each follower, the global optimal position is updated:

[0063] gbest=arg max(f(X i ))

[0064] Where: gbest is the global optimal solution; arg max(f(X i )) means finding the fitness function f(X i ) maximizes the individual X ilocation.

[0065] Preferably, in step 5, the Bootstrap resampling method is used to perform a random sampling on the residual data set {x1, x2, ..., x n} is extracted n times with replacement to generate a new Bootstrap sample set.

[0066]

[0067] For each Bootstrap sample x * , calculate the estimated Usually the mean:

[0068]

[0069] For all Bootstrap sample estimators, calculate their variance:

[0070]

[0071] Where: is the mean of all Bootstrap sample estimators; B is the number of samples.

[0072] Construct the upper and lower bounds of the confidence interval with a confidence level of 1-α:

[0073]

[0074]

[0075] Where: is the upper limit of the confidence interval; is the lower limit of the confidence interval; Z α / 2 It is the upper quantile of the standard normal distribution. The evaluation index of the model adopts interval coverage, interval standardized average width and coverage width comprehensive criteria as the evaluation index.

[0076] Compared with the prior art, the present invention has the following beneficial effects:

[0077] 1. The present invention organizes the historical monthly scale data of the selected water-wind-solar complementary system, including hydropower time series data, wind power time series data, photovoltaic power time series data, total power time series data and various climate system indexes.

[0078] 2. The present invention considers multiple lags and uses the maximum mutual information coefficient method to calculate the correlation between the climate system index and hydropower, wind power, photovoltaic power, and total power, and uses this as a basis to screen the prediction factors corresponding to each power, and considers the complementarity between hydropower, wind power, and photovoltaic power. Construct point forecast and interval forecast models based on deep learning models.

[0079] 3. The present invention considers two forecasting scenarios, namely, forecasting each power individually and then forecasting the total power cumulatively, and forecasting the total power jointly after accumulating each power. Cumulative forecasting and joint forecasting models are constructed based on long short-term memory network models. An optimization model is constructed with the maximum Nash efficiency coefficient and the maximum forecast error complementarity rate CRFE as objective functions. The sparrow optimization algorithm SSA is used to perform multi-objective optimization and solve the problem, and the parameters in LSTM (learning rate, regularization coefficient, number of hidden nodes) are optimized.

[0080] 4. The present invention adopts a non-parametric Bootstrap resampling method to resample from the residual data set between the point forecast prediction value and the true value to generate multiple new sample data sets, and superimposes the residuals obtained by these resampling on the point forecast value to generate multiple sets of new prediction results. Then, the upper and lower limits of the prediction interval are calculated based on these new prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 This is a flow chart of the multi-objective interval forecasting method for water, wind and solar power taking into account the complementarity of forecast errors of the present invention;

[0082] Figure 2 This is a diagram showing the hydropower power point prediction results in the hydro-wind-solar multi-energy complementary system of the present invention;

[0083] Figure 3 This is a wind power point prediction result diagram in the water-wind-solar multi-energy complementary system of the present invention;

[0084] Figure 4 This is a result diagram of photovoltaic power point prediction in the water-wind-solar multi-energy complementary system of the present invention;

[0085] Figure 5 This is a diagram showing the total power point prediction results in the water-wind-solar multi-energy complementary system of the present invention;

[0086] Figure 6 This is a diagram showing the hydropower power interval forecast result in the water-wind-solar multi-energy complementary system of the present invention;

[0087] Figure 7 This is a diagram showing the wind power range forecast result in the water-wind-solar multi-energy complementary system of the present invention;

[0088] Figure 8 This is a graph showing the photovoltaic power range prediction results in the water-wind-solar multi-energy complementary system of the present invention;

[0089] Fig. 9 This is a diagram of the total power range prediction results in the water-wind-solar multi-energy complementary system of the present invention. DETAILED DESCRIPTION

[0090] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0091] A multi-objective interval forecasting method for hydropower, wind power and solar power considering the complementarity of forecast errors comprises the following steps:

[0092] Step 1: Organize the historical monthly data of the selected water-wind-solar complementary system, including hydropower time series data, wind power time series data, photovoltaic power time series data, total power time series data and various climate system indexes;

[0093] Step 2: Considering the time lag effect of the prediction factor on the predicted power, the maximum mutual information coefficient method is used to calculate the correlation between each climate system index and each power, and based on this, the prediction factor corresponding to each power is selected from a large number of climate system indices;

[0094] Step 3: Analyze the complementarity between hydropower, wind power and photovoltaic power in the hydro-wind-solar multi-energy complementary system;

[0095] Step 4: construct various power forecasting models based on the long short-term memory network model, construct an optimization model with the maximum Nash efficiency coefficient and the maximum prediction error complementary rate CRFE as the objective function, use the sparrow optimization algorithm SSA to solve, optimize the parameters (learning rate, regularization coefficient, number of hidden nodes) in the LSTM model, and make a point forecast of the system power;

[0096] Step 5: Use the Bootstrap resampling method to calculate the residuals between the point prediction values ​​and the actual values ​​generated by the LSTM model for each power, resample the residuals to generate multiple sets of new residual data, and superimpose these resampled residuals on the point prediction values ​​to generate multiple sets of new prediction results. Then, calculate the upper and lower limits of the prediction interval based on these new prediction results to achieve interval prediction. The evaluation indicators of the model use interval coverage and interval standardized average width, and the comprehensive criterion of coverage width as evaluation indicators.

[0097] In this implementation, in step 2, the maximum mutual information coefficient is an index based on mutual information to measure the correlation between two variables, which can measure the amount of information contained in one random variable about another random variable. The calculation method is as follows:

[0098] Divide the scatter plot formed by the power prediction object np and the climate system index (or power) X in the dataset into grids, and calculate the mutual information value of each grid:

[0099]

[0100] where p(x) and p(n p) are the marginal probability density functions of the two variables; p(x,n p ) is the probability density function of the two variables,

[0101] Normalize the maximum mutual information value I*(X,N p ) in the grid:

[0102]

[0103] where a and b are the number of rows and columns of the grid division respectively; M(D) a,b is the normalized maximum mutual information value under the fixed grid division.

[0104] Perform grid divisions for different schemes and repeat the above steps to obtain all M(D) a,b The maximum value is the maximum mutual information coefficient MIC:

[0105]

[0106] where: ab < B(T) is the constraint condition; B(T) is usually taken as the 0.6th power of the dataset.

[0107] Adopt lag correlation analysis to determine the influence degree of the climate system index on each power under different lag periods. Lag analysis helps to understand the response speed and amplitude of one variable after another variable changes, which is of great significance for determining the optimal time step of the prediction factor and the prediction period of the prediction object. Analyze the lag influence of the climate system prediction factor on each power. The specific steps are to take the lag periods as 0 - 24 months (monthly scale) respectively, calculate the MIC values between each power and its climate system index, summarize the results, and draw a heat map of the lag correlation analysis between the climate system prediction factor and each power of the system to analyze the lag influence of the climate system prediction factor on each power, and select the climate system index with the most significant correlation and the corresponding lag time.

[0108] In step 2, through correlation analysis, the climate system prediction factors of the power of each single type of power station in the system are shown in Table 1.

[0109] Table 1 System power climate system prediction factors for each single type of power station

[0110]

[0111] In this embodiment, in step 3, the complementarity of hydropower, wind power and photovoltaic power is quantified, and the MIC method is used to calculate the mutual information index to quantify the complementarity. According to the analysis results, the complementarity between different energy sources is evaluated, and the mutual information is used to measure the nonlinear complementarity.

[0112] In this embodiment, in step 4, a joint forecast model and a cumulative forecast model of each power are constructed based on the LSTM model. For the joint forecast model, the climate system index corresponding to the power to be forecasted is used as input, and the power to be forecasted is used as output; for the cumulative forecast model, the climate system index corresponding to each power is used as input, and the power to be forecasted is used as output. After the three forecast models are established, the forecast results are accumulated to obtain the output result of the cumulative forecast model of the total power.

[0113] The principle of the LSTM model is as follows: LSTM (Long Short-Term Memory) is a special recurrent neural network (RNN) that introduces a gating mechanism to solve the gradient vanishing and gradient exploding problems that occur when traditional RNN processes long time series. The basic unit of the LSTM network includes three key gates: input gate, forget gate, and output gate. These gate mechanisms make LSTM particularly good at capturing long-term dependencies in time series. The LSTM model formula is as follows:

[0114] Input Gate:

[0115] i t =σ(W i ·[h t-1 ,X t ]+b i )

[0116] Where: i t is the input gate vector, which determines which new information will be stored in the cell state; σ is the Sigmoid activation function, which limits the output value between 0 and 1; W i is the weight matrix of the input gate; h t-1 is the hidden state of the previous time step; X t is the input vector of the current time step (including climate system indices of multiple time steps); b i Bias vector for the input gate.

[0117] Forget Gate:

[0118] f t =σ(W f ·[h t-1 ,X t ]+b f )

[0119] Where: f t is the forget gate vector, which determines which information is forgotten from the cell state; W f is the weight matrix of the forget gate; b f is the bias vector of the forget gate.

[0120] Candidate memory cell status:

[0121]

[0122] Where: is the candidate memory cell state, indicating a new candidate memory; tanh is the hyperbolic tangent activation function, which limits the output value to between -1 and 1; W C is the weight matrix of the candidate memory cell state; b C is the bias vector of the candidate memory cell state.

[0123] Memory cell status:

[0124]

[0125] Where: C t is the memory cell state at the current time step; f t is the forget gate vector; * is element-by-element multiplication, which does not need to meet the dimension requirements of matrix multiplication during calculation; C t-1 is the memory cell state at the previous time step.

[0126] Output Gate:

[0127] o t =σ(W o ·[h t-1 ,X t ]+b o )

[0128] Where: t is the output gate vector, which determines which parts of the cell state are output; W o is the weight matrix of the output gate; b o is the bias vector of the output gate.

[0129] Hidden state:

[0130] h t =o t *tanh(C t )

[0131] Where: h t is the hidden state at the current time step.

[0132] Construct an LSTM key parameter optimization model (learning rate, regularization coefficient, number of hidden nodes), with the maximum Nash-Sutcliffe efficiency coefficient (NSE) and the maximum complementary rate of forecast error CRFE (Complementary Rate of Forecast Error) as the objective function, and the parameter value range as the constraint condition. The Nash efficiency coefficient calculation formula is as follows:

[0133]

[0134] Where: is the point prediction value of the corresponding power in the tth period; N t is the actual value of the corresponding power in the tth period; T is the total number of time periods of the sample data; is the mean value of power.

[0135] The calculation formula of forecast error complementarity rate is as follows:

[0136]

[0137] Where: M is the total number of samples; M p is the number of hydropower, wind and solar power stations; is the jth predicted value of the ith sample; N i,j is the jth actual value of the ith sample. The forecast error complementarity rate is obtained by normalizing the difference between the numerator and the denominator to a value between 0 and 1. The closer CRFE is to 1, the smaller the model prediction error is, that is, the better the complementarity between different powers is; the closer it is to 0, the larger the model prediction error is, that is, the poorer the complementarity between different powers is.

[0138] The sparrow optimization algorithm SSA multi-objective function is used to solve the problem, optimize the parameters in LSTM (learning rate, regularization coefficient, number of hidden nodes), and predict the system power. The sparrow optimization algorithm (SSA) is used to solve the single-objective optimization problem, and the weighted summation method is used to merge the two objectives into one objective function. The specific steps are as follows:

[0139] Set the population size N and the maximum number of iterations T. Initialize the population position X:

[0140]

[0141] Where: X ij is the position of the i-th sparrow in the j-th dimension; and is the lower and upper bounds of the jth dimension; rand() generates a random number between 0 and 1; d is the time step;

[0142] Calculate the fitness of each sparrow:

[0143] f(X i )=w1·NSE(X i )+w2·CRFE(X i )

[0144] Where: w1 and w2 are the weight coefficients of NSE and CRFE; NSE(X i ) is the NSE value corresponding to the i-th sparrow position; CRFE(X i ) is the CRFE value corresponding to the i-th sparrow position.

[0145] Sort by fitness value and select the top N L The sparrow acts as the leader and the rest as followers. Leader position update:

[0146]

[0147] Where: is the position of the i-th leader at the t+1th iteration; is the position of the i-th leader at the t-th iteration; λ is a random number.

[0148] Follower Location Updates:

[0149]

[0150] is the position of the i-th follower at the t+1th iteration; is the position of the ith follower at the tth iteration; β is a random number; The best position for the navigator.

[0151] By comparing the fitness of each follower, the global optimal position is updated:

[0152] gbest=arg max(f(X i ))

[0153] Where: gbest is the global optimal solution; arg max(f(X i )) means finding the fitness function f(X i ) maximizes the individual X i location.

[0154] Table 2 shows the error indicators of the prediction results of each power point in the system

[0155]

[0156] In this embodiment, in step 5, the Bootstrap resampling method is used to perform a random sampling on the residual data set {x1, x2, ..., x n} is extracted n times with replacement to generate a new Bootstrap sample set.

[0157]

[0158] For each Bootstrap sample x * , calculate the estimated Usually the mean:

[0159]

[0160] For all Bootstrap sample estimators, calculate their variance:

[0161]

[0162] Where: is the mean of all Bootstrap sample estimators; B is the number of samples.

[0163] Construct the upper and lower bounds of the confidence interval with a confidence level of 1-α:

[0164]

[0165]

[0166] Where: is the upper limit of the confidence interval; is the lower limit of the confidence interval; Z α / 2 It is the upper quantile of the standard normal distribution. The evaluation index of the model adopts interval coverage, interval standardized average width and coverage width comprehensive criteria as the evaluation index.

[0167] Table 3 shows the comparison of the prediction effect of each power interval of the system

[0168]

[0169] In summary: the present invention organizes the historical monthly-scale data of the selected water-wind-solar complementary system, including hydropower time series data, wind power time series data, photovoltaic power time series data, total power time series data and various climate system indices, considers multiple lags, and adopts the maximum mutual information coefficient method to calculate the correlation between the climate system index and hydropower, wind power, photovoltaic power, and total power, and uses this as a basis to screen the prediction factors corresponding to each power, and considers the complementarity between hydropower, wind power, and photovoltaic power. Based on the deep learning model, point forecast and interval forecast models are constructed. Two forecast scenarios are considered, namely, forecasting each power individually and then forecasting the total power cumulatively, and forecasting the total power jointly after accumulating each power. Cumulative forecast and joint forecast models are constructed based on the long short-term memory network model. The optimization model is constructed with the maximum Nash efficiency coefficient and the maximum forecast error complementarity rate CRFE as the objective function. The sparrow optimization algorithm SSA is used for multi-objective optimization solution, and the parameters in LSTM (learning rate, regularization coefficient, number of hidden nodes) are optimized. The non-parametric Bootstrap resampling method is used to resample the residuals between the point prediction value and the actual value generated by each power to generate multiple sets of new residual data. The residuals obtained by these resampling are superimposed on the point prediction value to generate multiple sets of new prediction results. Then, according to the distribution of these new prediction sequences, the upper and lower limits of the prediction interval are calculated.

[0170] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0171] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-objective interval forecasting method for hydropower, wind power and solar power considering the complementarity of forecast errors, characterized by: The method includes the following steps: Step 1: Organize the historical monthly-scale data of the selected hydropower-wind-solar complementary system, including hydropower power time series data, wind power time series data, photovoltaic power time series data, total power time series data, and various climate system indices; Step 2: Considering the time lag effect of the prediction factors on the predicted power, use the maximum mutual information coefficient method to calculate the correlation between each climate system index and each power, and based on this, screen the prediction factors corresponding to each power from numerous climate system indices; Step 3: Analyze the complementarity among hydropower power, wind power, and photovoltaic power in the hydropower-wind-solar multi-energy complementary system; Step 4: Based on the long short-term memory network model, construct each power prediction model. Taking the maximum Nash efficiency coefficient and the maximum prediction error complementary rate CRFE as the objective function to construct an optimization model, and use the sparrow search algorithm SSA to solve it. Optimize the parameters (learning rate, regularization coefficient, number of hidden nodes) in the LSTM model to perform point prediction on the system power; Step 5: Adopt the Bootstrap resampling method. Calculate the residuals between the point prediction values generated by the LSTM model for each power and the actual values, resample the residuals to generate multiple groups of new residual data, superimpose these resampled residuals on the point prediction values to generate multiple groups of new prediction results. Then, calculate the upper and lower limits of the prediction interval according to these new prediction results, so as to achieve interval prediction. The evaluation indicators of the model are the interval coverage rate and the interval standardized average width, and the coverage width comprehensive criterion is used as the evaluation indicator.

2. The multi-objective interval forecasting method for water, wind and solar power considering forecast error complementarity according to claim 1 is characterized by: In Step 2, the maximum mutual information coefficient is an index based on mutual information to measure the correlation between two variables, and it can measure the amount of information contained in one random variable about another random variable. The calculation method is as follows: Perform grid division on the scatter plot formed by the power prediction object np and the climate system index (or power) X in the dataset, and calculate the mutual information value of each grid: In the formula, p(x) and p(n p) is the marginal probability density function of the two variables; p(x,n p ) is the probability density function of the two variables, The maximum mutual information value I*(X,N p )Normalization: Where: a and b are the number of rows and columns of the grid respectively; M(D) a,b is the normalized maximum mutual information value under fixed grid partitioning. Perform grid division with different schemes and repeat the above steps to obtain all M(D) a,b The maximum value is the maximum mutual information coefficient MIC: Where: ab < B(T) is the constraint condition; B(T) is usually taken as the 0.6th power of the dataset.

3. The multi-objective interval forecasting method for water, wind and solar power considering forecast error complementarity according to claim 1 is characterized by: In Step 3, to quantify the complementarity of hydropower power, wind power, and photovoltaic power, use the MIC method to calculate the mutual information index to quantify the complementarity. According to the analysis results, evaluate the complementarity between different energy sources, and use mutual information to measure the non-linear complementarity.

4. The multi-objective interval forecasting method for water, wind and solar power considering forecast error complementarity according to claim 1 is characterized by: In Step 4, based on the LSTM model, construct the joint prediction model and the cumulative prediction model for each power. For the joint prediction model, use the climate system index corresponding to the power to be predicted as the input and the power to be predicted as the output; For the cumulative prediction model, use the climate system index corresponding to each power as the input and the power to be predicted as the output respectively. After establishing three prediction models, accumulate the prediction results to obtain the output result of the cumulative prediction model of the total power. The principle of the LSTM model is as follows: LSTM (Long Short-Term Memory) is a special recurrent neural network (RNN) that introduces a gating mechanism to solve the gradient vanishing and gradient exploding problems that occur when traditional RNN processes long time series. The basic unit of the LSTM network includes three key gates: input gate, forget gate, and output gate. These gate mechanisms make LSTM particularly good at capturing long-term dependencies in time series. The LSTM model formula is as follows: Input Gate: i t =σ(W i ·[h t-1 ,X t ]+b i ) Where: i t is the input gate vector, which determines which new information will be stored in the cell state; σ is the Sigmoid activation function, which limits the output value between 0 and 1; W i is the weight matrix of the input gate; h t-1 is the hidden state of the previous time step; X t is the input vector of the current time step (including climate system indices of multiple time steps); b i Bias vector for the input gate. Forget Gate: f t =σ(W f ·[h t-1 ,X t ]+b f ) Where: f t is the forget gate vector, which determines which information is forgotten from the cell state; W f is the weight matrix of the forget gate; b f is the bias vector of the forget gate. Candidate memory cell status: Where: is the candidate memory cell state, indicating a new candidate memory; tanh is the hyperbolic tangent activation function, which limits the output value to between -1 and 1; W C is the weight matrix of the candidate memory cell state; b C is the bias vector of the candidate memory cell state. Memory cell status: Where: C t is the memory cell state at the current time step; f t is the forget gate vector; * is element-by-element multiplication, which does not need to meet the dimension requirements of matrix multiplication during calculation; C t-1 is the memory cell state at the previous time step. Output Gate: the t =σ(W o ·[h t-1 ,X t ]+b o ) In the formula: o t is the output gate vector, which determines which parts of the cell state are output; W o is the weight matrix of the output gate; b o is the bias vector of the output gate. Hidden state: h t =o t *tanh(C t ) Where: h t is the hidden state at the current time step. Construct an LSTM key parameter optimization model (learning rate, regularization coefficient, number of hidden nodes), with the maximum Nash-Sutcliffe efficiency coefficient (NSE) and the maximum complementary rate of forecast error CRFE (Complementary Rate of Forecast Error) as the objective function, and the parameter value range as the constraint condition. The Nash efficiency coefficient calculation formula is as follows: Where: is the point prediction value of the corresponding power in the tth period; N t is the actual value of the corresponding power in the tth period; T is the total number of time periods of the sample data; is the mean value of power. The calculation formula of forecast error complementarity rate is as follows: Where: M is the total number of samples; M p is the number of hydropower, wind and solar power stations; is the jth predicted value of the ith sample; N i,j is the jth actual value of the ith sample. The forecast error complementarity rate is obtained by normalizing the difference between the numerator and the denominator to a value between 0 and 1. The closer CRFE is to 1, the smaller the model prediction error is, that is, the better the complementarity between different powers is; the closer it is to 0, the larger the model prediction error is, that is, the poorer the complementarity between different powers is. The sparrow optimization algorithm SSA multi-objective function is used to solve the problem, optimize the parameters in LSTM (learning rate, regularization coefficient, number of hidden nodes), and predict the system power. The sparrow optimization algorithm (SSA) is used to solve the single-objective optimization problem, and the weighted summation method is used to merge the two objectives into one objective function. The specific steps are as follows: Set the population size N and the maximum number of iterations T. Initialize the population position X: Where: X ij is the position of the i-th sparrow in the j-th dimension; and is the lower and upper bounds of the jth dimension; rand() generates a random number between 0 and 1; d is the time step; Calculate the fitness of each sparrow: f(X i )=w1·NSE(X i )+w2·CRFE(X i ) Where: w1 and w2 are the weight coefficients of NSE and CRFE; NSE(X i ) is the NSE value corresponding to the i-th sparrow position; CRFE(X i ) is the CRFE value corresponding to the i-th sparrow position. Sort by fitness value and select the top N L The sparrow acts as the leader and the rest as followers. Leader position update: Where: is the position of the i-th leader at the t+1th iteration; is the position of the i-th leader at the t-th iteration; λ is a random number. Follower Location Updates: is the position of the i-th follower at the t+1th iteration; is the position of the ith follower at the tth iteration; β is a random number; The best position for the navigator. By comparing the fitness of each follower, the global optimal position is updated: gbest=argmax(f(X i )) Where: gbest is the global optimal solution; argmax(f(X i )) means finding the fitness function f(X i ) maximizes the individual X i location.

5. The multi-objective interval forecasting method for water, wind and solar power considering forecast error complementarity according to claim 1 is characterized by: In step 5, the Bootstrap resampling method is used to resample the residual data set {x1,x2,...,x n } is extracted n times with replacement to generate a new Bootstrap sample set. For each Bootstrap sample x * , calculate the estimated Usually the mean: For all Bootstrap sample estimators, calculate their variance: Where: is the mean of all Bootstrap sample estimators; B is the number of samples. Construct the upper and lower bounds of the confidence interval with a confidence level of 1-α: Where: is the upper limit of the confidence interval; is the lower limit of the confidence interval; Z α / 2 It is the upper quantile of the standard normal distribution. The evaluation index of the model adopts interval coverage, interval standardized average width and coverage width comprehensive criteria as the evaluation index.

Images