Icing rate prediction method based on improved particle swarm back propagation neural network

Abstract

The invention relates to an icing rate prediction method based on an improved particle swarm back propagation neural network, and relates to the field of power transmission line ice disaster prevention and control strategy research, and the method comprises the steps: building an icing growth mechanism model of a power transmission line in a low-temperature rain and snow freezing environment; changing multiple environmental parameters through numerical simulation, and performing multi-factor analysis by using an icing growth mechanism model to obtain icing working condition data such as multiple scenes, multiple moments, multiple weather and multiple terrains; four environmental characteristic values including the liquid water content, the environment temperature, the water drop diameter and the wind speed in the air are used as the input quantity of the network, the icing rate of the power transmission line is used as the output quantity of the network, and therefore a BPNN prediction model is constructed; meanwhile, the weight and the threshold value of the BPNN are optimized through IPSO, and accurate prediction of the icing rate of the power transmission line is achieved. The method is simple and feasible, the algorithm optimization speed is high, the model prediction accuracy is high, and reference is provided for power transmission line icing prediction and early warning work.

Description

Technical Field

[0001] The present invention relates to the research field of ice disaster prevention and control strategies for power transmission lines, and in particular to an ice coverage rate prediction method based on an improved particle swarm back propagation neural network. Background Art

[0002] Icing on transmission lines can cause power system disasters such as line breakage and line shaking. In more serious cases, tower collapse may occur, leading to large-scale power outages and endangering the safety of the power system.

[0003] Commonly used in the prediction of transmission line icing are mathematical and physical relationship models and intelligent algorithms based on neural networks. Mathematical and physical models require high-precision meteorological parameters, which are difficult to achieve in practical applications. Intelligent algorithms based on artificial neural networks are widely used in the field of icing rate prediction due to their unique advantages. BPNN is a pattern recognition method with self-organization, self-adaptation and self-learning capabilities. It can realize nonlinear mapping between meteorological parameters and line icing rate, and is suitable for solving difficult problems such as line icing rate prediction. However, BPNN is sensitive to weights, and artificial selection of initial weights and thresholds can easily cause the fitting results to fall into local extreme values. In recent years, researchers have combined the PSO algorithm with the BPNN algorithm to predict the icing rate, realizing automatic optimization, but the PSO algorithm is prone to fall into local optimization, which directly affects the performance of the algorithm and thus affects the prediction accuracy. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies of the prior art and propose a method for predicting icing rate based on an improved particle swarm back propagation neural network. The method takes multiple environmental factors of the transmission line as network input and the icing rate as network output, and combines an improved nonlinear adaptive inertia weighted particle swarm optimization algorithm with BPNN to achieve high-precision and rapid prediction of the icing rate of the transmission line.

[0005] To achieve the purpose of the present invention, the technical solution is as follows:

[0006] A method for predicting ice coverage rate based on improved particle swarm back propagation neural network is proposed. The specific implementation steps are as follows:

[0007] Step 1: Establish an ice growth mechanism model for transmission lines in a low-temperature rain, snow and freezing environment;

[0008] Step 2: Change multiple environmental parameters through numerical simulation, use the ice growth mechanism model to perform multi-factor analysis, and obtain ice condition data for multiple scenes, multiple times, multiple weather conditions, and multiple terrains;

[0009] Step 3: Take the four environmental characteristic values ​​of liquid water content, ambient temperature, water droplet diameter, and wind speed as the input of the network, and the transmission line icing rate as the output of the network to build a BPNN prediction model;

[0010] Step 4: Design nonlinear adaptive inertia weights for the BPNN prediction model, improve the particle swarm optimization algorithm, strengthen the global search capability and refine the local search capability;

[0011] Step 5: Implement the IPSO-BPNN transmission line icing rate prediction method.

[0012] Furthermore, the ice growth mechanism model of the transmission line in the low temperature rain, snow and freezing environment in step 1 is:

[0013] dm i =α 1i α 3i w i U i D i LdT i (1)

[0014] Among them, α 1i , α 3i , U i , D i 、m i They represent the collision rate α1, freezing coefficient α3, liquid water content w, wind speed U, ice diameter D and ice mass m of the transmission line in the i-th time step, d represents the differential, dm i represents the differential of the ice mass m on the line in the i-th time step, dt i It represents the differential of time t in the i-th time step, i represents the time step, and L represents the length of the transmission line.

[0015] First, the collision rate coefficient α1 is calculated by analyzing the trajectory of the water droplet and expressing it with two dimensionless functions. The values ​​of the dimensionless coefficients are shown in the following table.

[0016]

[0017] The overall collision rate α1 is used to determine the total ice growth rate, which refers to the amount of water droplets that actually collide on the windward surface of the entire transmission line and the amount of water droplets that may collide if the water droplet trajectory is assumed to be undeflected, and is defined as:

[0018]

[0019] Where y is the trajectory of water droplets, and s is the windward surface area of ​​the transmission line.

[0020] To calculate the collision rate, it is necessary to know the trajectory of the water droplets in the airflow, and the trajectory of the water droplets is determined by the motion equation of the water droplets.

[0021] When the airflow reaches the surface of the cylinder, the forces on the water droplets are: the relative speed generated after the airflow trajectory deviates from the water drop trajectory Viscous drag due to the viscosity of the air The water drop's own gravity Buoyancy in air currents The drag caused by the pressure difference between the front and rear of the water drop and apparent mass force → indicates a vector. According to Newton's second law, the force on a water droplet in air is for:

[0022]

[0023] Where: m d is the mass of the water drop, unit: kg; is the acceleration of the water droplet, unit: m / s 2 .

[0024] The gravity, buoyancy, pressure difference resistance and apparent mass force acting on the water droplet are all smaller than the viscous resistance, so they can be ignored without causing too much error.

[0025]

[0026] In the formula, R d is the radius of the water droplet, unit: m; ρ d , a are the densities of water droplets and gas, respectively, in kg / m 3 ; S d is the projected area of ​​the water drop, Unit: m 2 ; Respectively represent the speed of airflow and water droplets, unit: m / s; C D is the air damping coefficient, and g is the acceleration due to gravity.

[0027] And because the Reynolds number Re of the relative motion of water droplets in the airflow is:

[0028]

[0029] Where μ is the kinematic viscosity of air, unit: m / s 2 .

[0030] From equations (4) and (5), the differential equation of water drop motion is:

[0031]

[0032] Write (6) in dimensionless form:

[0033]

[0034] In the formula, represents the dimensionless velocity of airflow and water droplet, respectively, which are fractions based on the initial velocity U; τ is the dimensionless time, i.e., the time required for a water droplet moving at the airflow velocity U to advance the distance of the transmission line radius R; K is the Stokes number of the moving particle inertia:

[0035]

[0036] According to the above analysis, the collision rate of supercooled water droplets in the icing process is a function of wind speed U and water droplet diameter D. d and the transmission line diameter D. The water droplet collision rate can be expressed as two dimensionless parameters K and The function of K and They are defined as:

[0037]

[0038] The analytical calculation formula of the collision rate is obtained by performing nonlinear regression analysis:

[0039]

[0040] Secondly, the freezing coefficient α3 is calculated by analyzing the heat balance equation when the line is covered with ice, and substituting the parameters into it to solve the freezing coefficient.

[0041] The freezing coefficient is determined by the heat balance equation of the ice-covered surface, which is:

[0042] q f +q v +q k +q a +q n =q c +q e +q l +q s +q i +q r (12)

[0043] The heat flow terms in the heat balance equation are:

[0044] The latent heat released during the freezing of the α3 portion of the water droplet that collides with and remains on the wire surface is:

[0045] q f =2Rα1α3wU1L f (13)

[0046] Where α1 and α3 are the collision rate and freezing coefficient respectively; R is the radius of the ice-covered line, unit: m; w is the liquid water content, that is, the number of water droplets contained in each cubic meter of airflow, unit: g / m 3 ; U1 is the incoming flow velocity, unit: m / s; L f is the latent heat of melting of ice, which is generally taken as 3.35×10 5 J / kg.

[0047] Frictional heating of the conductor by air flow:

[0048] q v =2Rhr c U 2 / 2c a (14)

[0049] In the formula, r c is the local viscous heating recovery coefficient of the line cylindrical conductor surface, which is taken as 0.79; · represents the multiplication relationship; c a is the specific heat of air, which is 1014 J / kg·K; h is the convective heat transfer coefficient, in J / m 2 K:

[0050]

[0051] Where K a is the heat transfer coefficient of air, K a =0.0244; D i is the line diameter after ice cover, unit: m; N u is the Nusselt number, which can be expressed as:

[0052] N u =CRe n Pr 1 / 3 (16)

[0053] Where C and n are constants. For ice-covered environments, C = 0.683, n = 0.466; Re is the Reynolds number; Pr is the Prandtl number, which can be expressed as:

[0054]

[0055] Where D is the diameter of the ice-covered line, unit: m; d i is the ice thickness of the line, unit: m; v is the kinematic viscosity of air, v = 1.328 × 10 -5 ; a is the thermal diffusivity of air, a = 1.88 × 10 -5 .

[0056] Water drop collision kinetic energy:

[0057] q k =α1α2wRc3 (18)

[0058] Where c is the water droplet collision velocity, unit: m / s.

[0059] After the water droplets freeze into ice, they move from the freezing point T F (273.15K) to the equilibrium temperature T of the ice-covered surface s Heat released when:

[0060] q a =Dα1α2wUc(T F -T s ) (19)

[0061] For the wet growth process of conductor ice, due to T F =T s =273.15K, this item does not exist.

[0062] Energy gained by shortwave radiation q n , because ice usually occurs on foggy, rainy or cloudy days without direct sunlight, so this item is usually ignored;

[0063] Convective heat loss:

[0064] q c =2πRh(T s -T)(20)

[0065] Where h is the convective heat transfer coefficient of the ice-covered surface; R is the line radius; T is the ambient temperature; T s It is the surface temperature of ice cover when the ice cover is in dynamic equilibrium.

[0066] Latent heat loss from evaporation of liquid water or sublimation of ice:

[0067] q e =2πRχ(e(T s )-e(T))(21)

[0068] Where χ is the evaporation or sublimation coefficient, χ = 0.62hL e / (C a P), where L e =2.51×106 at T s The latent heat of evaporation or sublimation at temperature; k a is the thermal conductivity of the air medium; P is the air pressure; C a is the specific heat of air; e(T) is the saturated water vapor pressure of the water or ice surface on the ice-covered surface at temperature T, which can be calculated as follows:

[0069]

[0070] The supercooled water droplets that hit the surface of the transmission line rise from the ambient temperature T to the freezing temperature of the water droplets T F (273.15K) Heat absorbed:

[0071] q l =2Rα1α2Uwc w (T F -T)(23)

[0072] Where c w is the specific heat of water, c w =4.2×10 3 , unit: J / kg·K.

[0073] Heat lost to longwave radiation:

[0074] q s =8πRεσ R T 3 (T s -T)(24)

[0075] Where ε is the emissivity of the outer surface of the ice layer, ε = 0.95; σ R is the Stefan-Boltzman constant, which is 5.567×10 -8 .

[0076] Conduction heat loss:

[0077]

[0078] Where k is the thermal conductivity of the medium. For different heat conduction stages, k is the thermal conductivity of the response medium; A is the heat conduction area; is the temperature gradient in the normal direction of heat conduction; Δ is the change; R0 is the initial diameter of the ice-covered cylinder; R i is the diameter of the ice-covered cylinder. The heat conduction of the ice-covered conductor covered by ice is very small and can be ignored.

[0079] The heat taken away by the unfrozen supercooled water droplets leaving the ice surface:

[0080] q r =2Rα1α2wUc w (1-α3)(T s -T) (26)

[0081] Substituting each formula into (11), the freezing coefficient α3 can be obtained as follows:

[0082]

[0083] Furthermore, in step 2, multiple environmental parameters are changed through numerical simulation, and the ice growth mechanism model is used to perform multi-factor analysis to obtain ice condition data for multiple scenes, multiple times, multiple weather conditions, and multiple terrains:

[0084] Combined with the ice growth mechanism model of transmission lines in low-temperature rain, snow and freezing environment, four environmental parameters including liquid water content in the air, ambient temperature, water droplet diameter and wind speed are changed to analyze and calculate the line surface water droplet collision coefficient and surface water droplet freezing coefficient, and obtain icing condition data in multiple scenarios, multiple times, multiple weather conditions and multiple terrains.

[0085] Furthermore, in step 3, the four environmental characteristic values ​​of liquid water content, ambient temperature, water droplet diameter, and wind speed are used as the input of the network, and the transmission line icing rate is used as the output of the network to construct the BPNN prediction model:

[0086] The icing condition data was analyzed, and the four environmental characteristic values ​​of liquid water content, ambient temperature, water droplet diameter and wind speed were taken as the input of the network. The icing rate of the transmission line was taken as the output of the network to construct a BPNN prediction model.

[0087] Furthermore, in step 4, for the BPNN prediction model, a nonlinear adaptive inertia weight is designed, the particle swarm optimization algorithm is improved, the global search capability is enhanced, and the local search capability is refined as follows:

[0088] Description of related concepts and principles:

[0089] (1) BPNN structure

[0090] BPNN is a typical multi-layer forward neural network, which consists of input layer, hidden layer and output layer. Its learning process consists of two processes: forward propagation of information and back propagation of error. The input information from the outside world is calculated through the input layer, hidden layer and output layer to obtain the actual result of the network output; when the actual result does not match the expected result, the weights and thresholds of each layer are corrected by the error gradient descent method, and back propagated to the hidden layer and input layer. Through this continuous forward propagation and error gradient back propagation learning process, the weights and thresholds between each node are adjusted until the error between the network output and the actual output result reaches the set accuracy or the learning reaches a preset number of times.

[0091] The BPNN topology is as follows Figure 2 As shown, the input layer has m neural units, the hidden layer has p neural units, and the output layer has n neural units. According to the calculation formula of the neural unit output signal, the output of each neural unit in the hidden layer can be obtained as:

[0092]

[0093] In the formula, wij is the weight between input layer neural unit i and hidden layer neural unit j; b j is the threshold of the hidden layer neural unit j; net represents the network; f(·) is the transfer function of the hidden layer neural unit.

[0094] Similarly, the output of each neural unit in the output layer is:

[0095]

[0096] In the formula, w jk is the weight between hidden layer neural unit j and output layer neural unit k; θ k is the threshold of the output layer neural unit k; ψ(·) is the transfer function of the output layer neural unit.

[0097] (2) Particle Swarm Optimization Algorithm

[0098] PSO is a swarm optimization algorithm. In the particle swarm algorithm, the characteristics of each particle are represented by three indicators: position, speed, and fitness value. In each iteration, the particle tracks itself through two optimal values. The first is the optimal position of the fitness value calculated by the fitness function in the position passed by each particle, that is, the individual optimal value pbest. The other is the position corresponding to the optimal fitness value searched by all particles in the entire group, that is, the group optimal value gbest. Every time a particle updates a position, the corresponding fitness value is calculated according to the fitness function, and the position of the individual optimal value extreme value and the group optimal value extreme value is updated by comparing the new particle fitness value with the individual optimal value and the group optimal value fitness value until the fitness reaches the set requirements or the maximum number of iterations is reached, and the group optimal position is output as the optimal solution in the solution space.

[0099] Assume that the optimization space is a D-dimensional space with n particles and the position of the i-th particle is X i =(x i1 ,x i2 ,…,x iD ), the velocity of the particle is denoted by V i =(v i1 ,v i2 ,…v iD );The individual optimal value is P i =(p i1 ,p i2 ,…p iD );The global optimal value of the entire population is P g =(p g1 ,p g2 ,Lp gD ). During the evolution process, each particle updates its position and velocity, and the formula is as follows:

[0100]

[0101] Where d = 1, 2, ..., D; w is the inertia weight; k is the current iteration number; c1 and c2 are both constants greater than zero, called individual factors and social factors respectively; r1 and r2 are random numbers between [0, 1].

[0102] (3) Nonlinear adaptive inertia weight

[0103] The inertia weight w reflects the speed of the particle at the current iteration and the position inherited from the previous iteration. The larger the inertia weight w, the stronger the global search capability, the weaker the local search capability, and the higher the diversity of the population; and the smaller the inertia weight w, the stronger the local optimization capability, the weaker the global optimization capability, and the lower the population diversity. For highly nonlinear and complex problems such as the mapping of the icing rate of transmission lines to environmental parameters, the linear decreasing inertia weight cannot converge well. Therefore, the present invention adopts an improved nonlinear inertia weight. In the early stage of iteration, the nonlinear inertia weight has a larger value and a strong global search capability; in the later stage of iteration, a smaller inertia weight has a good local search capability.

[0104] Specifically, the following steps are included:

[0105] 1) IPSO weight adjustment

[0106]

[0107] In the formula, w s is the initial value of the inertia weight; w e is the inertia weight to reach the maximum number of iterations; t is the current number of iterations; T is the set maximum iteration coefficient; α is the adjustment coefficient.

[0108] 2) Optimization of BPNN

[0109] Aiming at the shortcomings of traditional BPNN, such as slow convergence and easy falling into local optimum, the mean square error function of BPNN is used as the fitness function of PSO algorithm, and adaptive nonlinear inertia weight is adopted to find the particle with the smallest fitness value in PSO algorithm, and the position vector of this particle is used as the weight and threshold of BPNN.

[0110] Furthermore, the transmission line icing rate prediction method of implementing IPSO-BPNN in step 5 is:

[0111] The implementation process of an ice coverage rate prediction method based on an improved particle swarm back propagation neural network is as follows: Figure 3 As shown, the following steps are mainly involved:

[0112] (1) Determine the BPNN topology structure, normalize the icing condition data, and form feature vectors with the normalized multi-environmental parameter eigenvalues ​​and the icing rate of the transmission line to form training set samples and test set samples to input into the neural network.

[0113] (2) Initialize the particle swarm parameters and set w s ,w e Population size, maximum number of iterations T, current number of iterations t, adjustment coefficient α, etc.

[0114] (3) Select the error function:

[0115]

[0116] is the fitness function of the particle swarm optimization algorithm. fitness is the error function, n is the number of training samples; k is the output value of the kth sample; t k is the target value of the kth sample. The fitness value is calculated by the fitness function to determine the advantage of the particle, and the individual optimal position of each particle is set to pbest; the optimal position experienced by the entire group is set to gbest.

[0117] (4) Update the position and velocity of each particle according to the particle position and velocity update formula, and reset the position of particles that are out of the search space.

[0118] (5) Calculate the fitness value of the updated particles and update the individual optimal position and group optimal position of each particle.

[0119] (6) Determine whether the termination condition is met. The termination condition is whether the maximum number of iterations is reached or the fitness value meets the set accuracy requirement. If the termination condition is met, stop searching, output the optimal position of the group, and execute (7); otherwise, return to (4) to continue searching.

[0120] (7) The optimal position of the group is converted into the weights and thresholds corresponding to the BPNN, and the network is trained. At the same time, the network weights and thresholds are updated. When the training requirements are met, the training is completed. The test set samples are input into the network to obtain the test results of the network.

[0121] Based on the construction of the transmission line icing model, the present invention proposes an icing rate prediction method based on an improved particle swarm back propagation neural network, which solves the problems of slow convergence speed, low recognition accuracy, and poor generalization performance of traditional BPNN, thereby making the transmission line icing rate prediction more reliable and effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0122] Figure 1It is a calculation flow chart of the ice growth mechanism model of the transmission line in the low temperature rain, snow and freezing environment;

[0123] Figure 2 It is the BPNN topology diagram;

[0124] Figure 3 It is the flow chart of IPSO-BPNN algorithm for transmission line icing rate prediction;

[0125] Figure 4 It is a transmission line icing rate prediction diagram based on IPSO-BPNN;

[0126] Figure 5 It is the best fitness change curve of IPSO-BPNN algorithm. DETAILED DESCRIPTION

[0127] Based on the extraction of ice data of power transmission lines in low-temperature rain, snow and freezing environments, the present invention uses the mean square error function of BPNN as the fitness function of the particle swarm optimization algorithm to solve the shortcomings of slow convergence and local optimality of traditional BPNN. It proposes an improved nonlinear adaptive inertia weighted particle swarm algorithm, finds the particle with the smallest fitness value in the PSO algorithm, and uses the vector position of this particle as the weight and threshold of BPNN to optimize BPNN, thereby improving the generalization ability of the neural network and realizing high-precision and rapid identification of ice thickness of power transmission lines. The specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0128] (1) Parameter setting of IPSO-BPNN algorithm for transmission line icing rate

[0129] 1) Initialize the particle swarm parameters and set w s =0.9,w e =0.4, the population size is 20, the maximum number of iterations is 500, the particle velocity range is [-1,1], the particle position range is [-5,5], and the particle dimension is set to the sum of the network weight and the threshold number.

[0130] 2) The IPSO-BPNN model structure is 4-5-1. The input layer is composed of four characteristic quantities, namely, liquid water content in the air, ambient temperature, water droplet diameter, and wind speed meteorological parameters. The number of nodes in the hidden layer is 5. The output layer outputs one characteristic quantity, namely, the ice coverage rate of the transmission line. The transfer function of the hidden layer is tansig. The transfer function of the output layer is purelin. The training function is trainlm, and the LM algorithm is used to train the neural network.

[0131] 3) The maximum number of network iterations is 1000, the learning rate is 0.01, the training accuracy is 0.001, and the ratio of training samples to test samples is 10:1.

[0132] 4) The adjustment coefficient α is set to 0.003. A BPNN optimized by nonlinear inertia weight IPSO is established, trained with training samples, and then 20 groups of test samples are input into the network.

[0133] (2) Identification results and analysis of the IPSO-BPNN algorithm for transmission line icing rate prediction

[0134] The prediction results of transmission line icing conditions are as follows: Figure 4 As shown, the optimal fitness change curve is as follows Figure 5 shown.

[0135] from Figure 5 It can be seen that the BPNN optimized by the particle swarm algorithm has high recognition accuracy. The BPNN optimized by the improved nonlinear adaptive inertia weight of the present invention has a fitness value that continues to decrease with the continuous iteration of the improved particle swarm algorithm, that is, the BPNN error continues to decrease and eventually tends to be stable. Moreover, the iterative optimization speed is faster, and the optimal fitness of the particles is lower, close to 0.7. The root mean square error of the IPSO-BPNN algorithm is close to 0. In view of the problem that PSO is prone to fall into local optimality, by adding learning factors and nonlinear changes in inertia weights to the PSO algorithm, the improved algorithm can have stronger global optimization capabilities and convergence speed. The superiority of the improved algorithm was verified by simulation of data.

[0136] The above examples are merely examples for clearly illustrating the present invention, rather than limitations on the implementation modes of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made on the basis of the above description to facilitate the understanding and implementation of the invention.

Claims

1. A method for predicting ice coverage rate based on improved particle swarm back propagation neural network, characterized in that: The steps include: Step 1: Establish an ice growth mechanism model for transmission lines in a low-temperature rain, snow and freezing environment; Step 2: Change multiple environmental parameters through numerical simulation, use the ice growth mechanism model to perform multi-factor analysis, and obtain multi-scenario, multi-time, multi-weather, and multi-terrain icing condition data; Step 3: Take the four environmental characteristic values ​​of liquid water content, ambient temperature, water droplet diameter, and wind speed as the input of the network, and the transmission line icing rate as the output of the network to build a BPNN prediction model; Step 4: Design nonlinear adaptive inertia weights for the BPNN prediction model, improve the particle swarm optimization algorithm, strengthen the global search capability and refine the local search capability; Step 5: Implement the IPSO-BPNN transmission line icing rate prediction method; The ice growth mechanism model of the transmission line in the low temperature rain, snow and freezing environment in step 1 is: dm i =a 1i a 3i w i U i D i Ldt i ; Among them, α 1i , α 3i 、w i , U i , D i 、m i are the collision rate (α1), freezing coefficient (α3), liquid water content (w), wind speed (U), ice diameter (D) and ice mass (m) of the transmission line in the i-th time step, d represents the differential, dm i represents the differential of the ice mass m on the line in the i-th time step, dt i represents the differential of time t in the i-th time step, i represents the time step, and L represents the length of the transmission line; In step 2, multiple environmental parameters are changed by numerical simulation, and a multi-factor analysis is performed using an ice growth mechanism model to obtain multi-scenario, multi-time, multi-weather, and multi-terrain icing condition data, specifically: Combined with the ice growth mechanism model of transmission lines in low-temperature rain, snow and freezing environment, the values ​​of multiple environmental parameters are changed, the line surface water droplet collision coefficient and surface water droplet freezing coefficient are analyzed and calculated, and the ice condition data of multiple scenes, multiple times, multiple weather conditions and multiple terrains are obtained; The step 4 specifically comprises the following steps: (1) The calculation formula of the improved particle swarm optimization algorithm after weight adjustment is as follows: In the formula, w s is the initial value of the inertia weight; w e is the inertia weight to reach the maximum number of iterations; t1 is the current number of iterations; T is the set maximum number of iterations; α is the adjustment coefficient; (2) BPNN optimization: The mean square error function of BPNN is used as the fitness function of the particle swarm algorithm. Adaptive nonlinear inertia weight is used to find the particle with the smallest fitness value in the PSO algorithm, and the position vector of this particle is used as the weight and threshold of BPNN. The IPSO-BPNN transmission line icing rate prediction method in step 5 is implemented as follows: (1) Determine the BPNN topology structure, normalize the input working condition data, and form a feature vector with the normalized eigenvalues ​​and ice coverage rate to form training set samples and test set samples to input into the neural network; (2) Initialize the particle swarm parameters and set w s 、w e , population size, maximum number of iterations T, current number of iterations t1, adjustment coefficient α; (3) Select the error function: is the fitness function of the particle swarm optimization algorithm; where E fitness is the error function, n is the number of training samples; k is the output value of the kth sample; t k is the target value of the kth sample; the fitness value is calculated by the fitness function to determine the quality of the particle, and the individual optimal position of each particle is set to pbest; the optimal position experienced by the entire group is set to gbest; (4) Update the position and velocity of each particle according to the particle position and velocity update formula, and reset the position of particles that are out of the search space; (5) Calculate the fitness value of the updated particles and update the individual optimal position and group optimal position of each particle; (6) Determine whether the termination condition is met, and the termination condition is whether the maximum number of iterations is reached or the fitness value meets the set accuracy requirement; if the termination condition is met, stop searching, output the optimal position of the group, and execute the following step (7); otherwise, return to step (4) to continue searching; (7) The optimal position of the group is converted into the weights and thresholds corresponding to the BPNN, the network is trained, and the network weights and thresholds are updated at the same time. When the training requirements are met, the training is completed; the test set samples are input into the network to obtain the test results of the network.

Images