Power load prediction method and system based on reinforcement learning and dynamic mode decomposition

Abstract

The invention discloses a power load prediction method and system based on reinforcement learning and dynamic mode decomposition. The method comprises the following steps: 1) data acquisition and preprocessing; 2) performing multi-stage adaptive noise filtering and modal decomposition, performing complementary set empirical mode decomposition on the preprocessed load data to obtain an initial IMF component set, performing noise reduction by using a dynamic wavelet threshold, performing reconstruction, and performing secondary decomposition by using VMD; 3) constructing and training a dynamic routing prediction network driven by reinforcement learning, and predicting data of different time scales by designing a dynamic routing prediction network based on deep reinforcement learning; and 4) performing uncertainty perception weighting and result fusion. The method improves the prediction precision, and guarantees the operation reliability of a power grid.

Description

Technical Field

[0001] The present invention relates to the field of smart grid technology, and in particular to a power load forecasting method and system based on reinforcement learning and dynamic mode decomposition. Background Art

[0002] With the continuous growth of global electricity demand, the stability and efficiency of power systems have become particularly important. As a key link in power system operation, power load forecasting plays an important role in optimizing the allocation of power resources, reducing operating costs, and improving power supply reliability and safety.

[0003] In the field of power load forecasting, although the combination of deep learning models and signal decomposition technology has made certain progress, the following key issues still exist: existing models are unable to dynamically adjust the calculation path according to the real-time characteristics of the load components, resulting in difficulty in balancing high-frequency noise and low-frequency trends.

[0004] Although traditional gating networks can achieve static branch selection, they lack an online optimization mechanism based on prediction results; in sudden load fluctuation scenarios, fixed weight fusion strategies are easily interfered by abnormal components.

[0005] Existing technologies rely on manual experience and are unable to adaptively suppress noise; traditional two-stage training leads to error accumulation, and the decomposition parameters and prediction model parameters are optimized in isolation, making global optimality difficult to ensure.

[0006] The above problems will lead to a decrease in prediction accuracy, which may cause problems such as incorrect power dispatching decisions and unreasonable backup capacity configuration, directly affecting the safety of power grid operation. Summary of the Invention

[0007] In response to the above problems, the present invention provides a power load forecasting method and system based on reinforcement learning and dynamic mode decomposition.

[0008] The technical solution of the present invention is: a power load forecasting method based on reinforcement learning and dynamic mode decomposition, comprising the following steps:

[0009] 1) Data acquisition and preprocessing: Acquire regional power load data and then preprocess the data;

[0010] 2) Multi-stage adaptive noise filtering and modal decomposition: The pre-processed load data is subjected to complementary set empirical mode decomposition to obtain an initial set of IMF components. Dynamic wavelet thresholding is then used for noise reduction. After reconstruction, VMD is used for secondary decomposition to obtain the IMF components.

[0011] 3) Based on the IMF component, a dynamic routing prediction network based on deep reinforcement learning is designed to predict data at different time scales;

[0012] 4) The uncertainty perception weight is integrated with the prediction result to obtain the final prediction result.

[0013] In step 1), the power load data of a certain area is obtained to form an original load data set, and the data set is repaired.

[0014] Step 2) includes:

[0015] 21) First, preselect the wavelet set and determine the maximum decomposition level L = log2(N)-3 according to the signal length, where N is the number of signal points;

[0016] 22) Dynamic calculation of local signal-to-noise ratio, the high-frequency components after complementary set empirical mode decomposition Divide into windows of length W, with adjacent windows overlapping by 50%. For the kth window S k (t) Calculate the signal-to-noise ratio (SNR) k :

[0017]

[0018] Where, is the signal after preliminary denoising, S k (t) is the original signal;

[0019] 23) Dynamic wavelet basis and threshold rule selection,

[0020] Wavelet basis selection strategy: If SNR k <5dB, select Symlet wavelet; if 5dB≤SNR k <15dB, select Daubechics wavelet; if SNR k ≥15dB, select Coiflet wavelet;

[0021] Threshold rule selection: If SNR k <5dB, using SURE threshold:

[0022]

[0023] Where λ is the adaptive threshold, σ is the noise standard deviation, and df(λ) is the effective degree of freedom;

[0024] If SNR k ≥15dB, using Minimax threshold:

[0025]

[0026] 24) Layered threshold processing, for each layer of wavelet coefficient d j(j=1,2,...,L) performs dynamic thresholding: select the threshold rule according to the SNR of the current window and calculate the inter-layer adaptive threshold λ j , and then soft thresholding:

[0027]

[0028] Where, d j (t) is the j-th layer wavelet decomposition coefficient, λ j is the dynamic threshold of the jth layer, Coefficients after thresholding;

[0029] Then perform local adaptive enhancement to enhance the amplitude of the retained wavelet coefficients. j (t)|>2λ j The coefficient of:

[0030]

[0031] Where α is the penalty factor;

[0032] 25) Dynamic reconstruction and post-processing: using dynamically selected wavelet basis to threshold the coefficients Reconstruct and obtain the denoised signal

[0033] The particle swarm algorithm is used to search for the optimal mode number K and penalty factor δ for the denoised signal. The objective function is:

[0034]

[0035] Perform VMD decomposition on the denoised components to obtain the final IMF set.

[0036] Step 3) includes: reinforcement learning driven dynamic routing prediction network construction and training, as follows:

[0037] 31) Time series feature extraction and state definition, for each IMF component IMF j , extract local variance Main frequency energy ratio E j , autocorrelation decay rate υ j , the average error MAE in the historical window history , construct the state vector:

[0038]

[0039] By emphasizing that the state vector contains dynamic temporal characteristics and historical errors, the context-awareness of the strategy is enhanced;

[0040] 32) Reinforcement Learning Strategy Network Design, Actor Network: Input State sj , output action probability distribution P(a j |s j ), the action space is {a BITCN , a BILSTM}; Critic network: input state s j , output value estimate V(s j );

[0041] 33) Dynamic routing decision and model reasoning, based on the action probability output by the Actor network, an ε-greedy strategy is used to select branches:

[0042]

[0043] Where ε is the exploration probability;

[0044] Execute action a j , if a j =a BITCN , the IMF j Enter a bidirectional TCN branch; if a j =a BILSTM , the IMF j Input bidirectional LSTM branch;

[0045] 34) Reward calculation and experience replay, introducing smoothing terms, and defining reward r j =-MAE j +τ·smooth term, where MAE j is the forecast error of the current IMF component, τ is the weight coefficient; the storage experience tuple (s j ,a j ,r j ,s′ j );

[0046] 35) Online optimization of the policy network, sampling batch data from the replay buffer and calculating the advantage function:

[0047] A(s j )=r j +γV(s j ′)-V(s j ) (9)

[0048] Where r j is the immediate reward, γ is the discount factor, V(s j ) is the value estimate of the current state, V(s j ′) is the value estimate of the next state;

[0049] Update the Actor Network:

[0050]

[0051] Where θ is the Actor network parameter, π θ For strategic networks;

[0052] Update the Critic network:

[0053] L critic =E[(r j +γV(s j ′)-V(s j )) 2 ] (11)

[0054] Where, L critic is the mean square error loss of the Critic network.

[0055] Step 4) includes:

[0056] 41) Confidence interval calculation, for each IMF j The prediction results Booststrap was used to sample 100 times and the 95% confidence interval was calculated. in, is the forecast result of the j-th IMF component, is the lower bound of the 95% confidence interval of the j-th IMF forecast result, is the upper limit of the 95% confidence interval of the j-th IMF forecast result;

[0057] 42) Dynamic weight allocation, weight is ε=1e-6,where w j is the fusion weight of the j-th IMF component, and ε is a minimum constant;

[0058] 43) Weighted accumulation, the final prediction result is in, is the final fused power forecast result, and m is the total number of IMF components.

[0059] The power load forecasting system based on reinforcement learning and dynamic mode decomposition includes:

[0060] Data acquisition and preprocessing module, used to acquire regional power load data and then preprocess the data;

[0061] The multi-stage adaptive noise filtering and modal decomposition module is used to perform complementary set empirical mode decomposition on the pre-processed load data to obtain the initial IMF component set; dynamic wavelet thresholding is then used for noise reduction, and after reconstruction, VMD is used for secondary decomposition to obtain the IMF components;

[0062] The dynamic prediction module is used to predict data of different time scales based on the IMF component by designing a dynamic routing prediction network based on deep reinforcement learning;

[0063] The fusion prediction module is used to fuse the uncertainty perception weighting with the prediction result to obtain the final prediction result.

[0064] In the work, this invention introduces reinforcement learning technology and proposes the use of a "decomposition-routing-prediction" full-process dynamic optimization framework. Reinforcement learning is used to drive the overall decomposition and prediction process. For each IMF component, multi-dimensional time series features such as local variance, main frequency energy ratio, and autocorrelation decay rate are extracted. The mixed state vector is constructed in combination with historical prediction errors to enhance the context perception ability of the reinforcement learning strategy; branch selection is modeled as a discrete action space, and the policy gradient algorithm is used to realize online adjustment of action probability to solve the problem of parameter solidification of traditional gated networks.

[0065] The present invention improves prediction accuracy and ensures reliability of power grid operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flow chart of the method of the present invention;

[0067] Figure 2 This is the CEEMD decomposition result diagram;

[0068] Figure 3 is the curve graph after denoising;

[0069] Figure 4 This is the VMD decomposition diagram;

[0070] Figure 5 This is the predicted result graph. DETAILED DESCRIPTION

[0071] The following describes the implementation scheme and specific operation process of the present invention in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following specific description.

[0072] like Figure 1 As shown, a power load forecasting method based on reinforcement learning and dynamic mode decomposition includes:

[0073] 1) Data acquisition and preprocessing: Obtain regional power load data and then preprocess the data. Publicly available power load data from a location in Australia (including real-time dry-bulb temperature, dew point temperature, humidity, and other data) is selected and preprocessed to provide high-quality input for subsequent decomposition.

[0074] 2) Multi-stage adaptive noise filtering and modal decomposition: Complementary Ensemble Empirical Mode Decomposition (CEEMDAN) is performed on the load data to obtain an initial set of IMF components. This is then denoised using a dynamic wavelet threshold. After reconstruction, the particle swarm algorithm (PSO) is used to optimize the penalty factor α and modal number K of variational mode decomposition (VMD) for secondary decomposition. The decomposed IMF components carry a signal-to-noise ratio label, providing key features for subsequent routing decisions.

[0075] 3) Construction and training of a dynamic routing prediction network driven by reinforcement learning. This approach, based on deep reinforcement learning, enables predictions for data at different time scales. By designing an actor-critic reinforcement learning policy network, prediction tasks are dynamically assigned based on characteristics such as the local variance and dominant frequency energy of the IMF components. High-SNR components are routed to a BiLSTM to capture long-term trends, while low-SNR components are routed to a BiTCN to handle localized mutations. While the traditional attention mechanism uses fixed weight fusion, it struggles to cope with sudden load changes. This approach, however, improves routing decision accuracy by constraining policy smoothness through a reward function.

[0076] 4) The uncertainty perception weights are integrated with the prediction results to obtain the final prediction result. This step uses Monte Carlo Dropout to perform bootstrap resampling on the prediction results, generate confidence intervals, calculate KL divergence to quantify uncertainty, and dynamically adjust the fusion weights. Traditional weighted averaging methods are sensitive to outliers, but this method uses probability distribution alignment constraints to automatically downweight prediction results during abnormal periods, reducing error fluctuations under abnormal interference.

[0077] In step 1): obtain the power load data of a certain area to form the original load data set. The power data, meteorological data, etc. can be obtained through open source websites at home and abroad or public data sets to form the original data. Since there are many outliers and missing values in the original data, this will have a negative impact on the performance of the prediction model, so the data set is repaired. Then divide the data set into training set and test set in an 8:2 ratio. By repairing outliers and filling missing values, data integrity is ensured and noise interference in model training is avoided. The 8:2 time series partitioning strategy is adopted to strictly isolate the training set and the test set to simulate the demand of "using historical data to predict the future" in real scenarios and prevent overfitting due to data leakage.

[0078] In step 2): the high-frequency components after CEEMDAN decomposition Perform dynamic wavelet threshold denoising, the result is as follows Figure 2The nine intrinsic mode functions (IMFs) generated by the decomposition show significant frequency band characteristics: IMF1-3 as high-frequency components capture instantaneous disturbances with an average period of 5-15 minutes; IMF4-6 as medium-term components reflect the daytime power consumption fluctuation pattern; IMF7-9 as low-frequency components reflect the long-term load trend.

[0079] Step 2) includes:

[0080] 21) First, preselect the wavelet set and determine the maximum decomposition level L = log2(N)-3 according to the signal length, where N is the number of signal points;

[0081] 22) Dynamic calculation of local signal-to-noise ratio, the high-frequency components after complementary set empirical mode decomposition Divide into windows of length W, with adjacent windows overlapping by 50%. For the kth window S k (t) Calculate the signal-to-noise ratio (SNR) k :

[0082]

[0083] Where, is the signal after preliminary denoising, S k (t) is the original signal;

[0084] 23) Dynamic wavelet basis and threshold rule selection,

[0085] Wavelet basis selection strategy: If SNR k <5dB, select Symlet wavelet; if 5dB≤SNR k <15dB, select Daubechics wavelet; if SNR k ≥15dB, select Coiflet wavelet;

[0086] Threshold rule selection: If SNR k <5dB, using SURE threshold:

[0087]

[0088] Where λ is the adaptive threshold, σ is the noise standard deviation, and df(λ) is the effective degree of freedom.

[0089] If SNR k ≥15dB, using Minimax threshold:

[0090]

[0091] 24) Layered threshold processing, for each layer of wavelet coefficient d j(j=1,2,...,L) performs dynamic thresholding: select the threshold rule according to the SNR of the current window and calculate the inter-layer adaptive threshold λ j , and then soft thresholding:

[0092]

[0093] Where, d j (t) is the j-th layer wavelet decomposition coefficient, λ j is the dynamic threshold of the jth layer, Coefficients after thresholding;

[0094] Then perform local adaptive enhancement to enhance the amplitude of the retained wavelet coefficients. j (t)|>2λ j The coefficient of:

[0095]

[0096] Where α is the penalty factor (α = 1.2);

[0097] 25) Dynamic reconstruction and post-processing, wavelet inverse: using dynamically selected wavelet basis to threshold the coefficients Reconstruct and obtain the denoised signal like Figure 3 As shown;

[0098] The particle swarm algorithm is used to search for the optimal mode number K and penalty factor δ for the denoised signal. The objective function is:

[0099]

[0100] Perform VMD decomposition on the denoised components to obtain the final IMF set. Figure 4 The results of variational mode decomposition (VMD) using particle swarm optimization are shown. Each mode exhibits a clear frequency band division: Mode 1 captures minute-by-minute fluctuations at the device level; Mode 2 analyzes midday peaks in electricity consumption; Mode 3 corresponds to daily load variations; and Modes 4-5 characterize weekly trend components.

[0101] Through signal decomposition and dynamic denoising, efficient load feature extraction is achieved, addressing the difficulty of balancing noise suppression and trend preservation in traditional methods. First, CEEMDAN decomposition splits the raw load data into high-frequency components (short-term fluctuations and noise) and low-frequency components (long-term trends and cycles), avoiding modal aliasing. Second, a dynamic wavelet threshold denoising algorithm is designed for the high-frequency components. The algorithm adaptively selects wavelet bases (Symlet for noise reduction and Coiflet for shape preservation) and thresholding rules (SURE threshold for optimized noise suppression and Minimax threshold for valid signal preservation) based on the local signal-to-noise ratio (SNR). Furthermore, a hierarchical thresholding and coefficient enhancement mechanism are introduced to simultaneously enhance key features while denoising. Finally, a particle swarm optimization algorithm is used to optimize VMD parameters, achieving precise frequency band segmentation through secondary decomposition. The key advantages of this step include: 1) noise suppression efficiency: dynamic wavelet denoising improves the signal-to-noise ratio of high-frequency components, making it particularly suitable for impulse noise in industrial loads; and 2) modal clarity: after secondary decomposition, the frequency band overlap of the IMF components is reduced, effectively separating harmonic interference caused by wind farm grid connection from actual load fluctuations.

[0102] In step 3), the construction and training steps of the reinforcement learning-driven dynamic routing prediction network are as follows: the Actor-Critic dual network framework is used to implement intelligent branch decision-making, in which the state perception layer constructs a multi-dimensional feature vector through four key indicators (local variance, main frequency energy ratio, autocorrelation decay rate, and historical window average error), which is input into the policy network. The policy network outputs the selection of the bidirectional temporal convolutional network (BiTCN) and the bidirectional long short-term memory network (BiLSTM), and achieves a balance between exploration and utilization through the ε-greedy strategy. The dynamic routing module inputs the IMF component into the specified model according to the decision. The reward feedback system calculates the compound reward value containing the smooth penalty term in real time, and realizes the online optimization of the network parameters through the policy gradient algorithm. The specific process is as follows:

[0103] 31) Time series feature extraction and state definition, for each IMF component IMF j , extract local variance Main frequency energy ratio E j , autocorrelation decay rate υ j , the average error MAE in the historical window history , construct the state vector:

[0104]

[0105] By emphasizing that the state vector contains dynamic temporal characteristics and historical errors, the context-awareness of the strategy is enhanced;

[0106] 32) Reinforcement Learning Strategy Network Design, Actor Network: Input State s j , output action probability distribution P(aj |s j ), the action space is {a BITCN , a BILSTM}; Critic network: input state s j , output value estimate V(s j );

[0107] 33) Dynamic routing decision and model reasoning, based on the action probability output by the Actor network, an ε-greedy strategy is used to select branches:

[0108]

[0109] Where ε is the exploration probability.

[0110] Execute action a j , if a j =a BITCN , the IMF j Enter a bidirectional TCN branch; if a j =a BILSTM , the IMF j Input bidirectional LSTM branch;

[0111] 34) Reward calculation and experience replay, introducing smoothing terms, and defining reward r j =-MAE j +τ·smooth term, where MAE j is the forecast error of the current IMF component, τ is the weight coefficient; the storage experience tuple (s j ,a j ,r j ,s′ j );

[0112] 35) Online optimization of the policy network, sampling batch data from the replay buffer and calculating the advantage function:

[0113] A(s j )=r j +γV(s j ′)-V(s j ) (9)

[0114] Where r j is the immediate reward, γ is the discount factor, balancing current and future rewards, V(s j ) is the value estimate of the current state, V(s j ′) is the value estimate of the next state.

[0115] Update the Actor Network:

[0116]

[0117] Where θ is the Actor network parameter, π θ For the policy network, output action probability distribution;

[0118] Update the Critic network:

[0119] L critic =E[(r j +γV(s j ′)-V(s j )) 2 ] (11)

[0120] Where, L critic is the mean square error loss of the Critic network.

[0121] The dynamic routing mechanism achieves dynamic matching of model structure and data characteristics through reinforcement learning, overcoming the limitations of traditional single models. First, the state vector is designed to capture the local variance (noise intensity) of the IMF component, the main frequency energy (periodicity), the autocorrelation decay rate (trend stability), and the historical prediction error, providing a multi-dimensional basis for routing decisions. Subsequently, the actor-critic network generates action probabilities based on the state vector (selecting either the BiTCN or BiLSTM branch). BiTCN excels at capturing short-term mutations (such as the start and stop of large motors), while BiLSTM is suitable for modeling long-term cycles (such as diurnal load fluctuations). A smoothing term is introduced into the reward function to penalize frequent switching and ensure policy stability.

[0122] In step 4), the uncertainty perception weighting and result fusion steps are as follows:

[0123] 41) Confidence interval calculation, for each IMF j The prediction results Booststrap was used to sample 100 times and the 95% confidence interval was calculated. in, is the forecast result of the j-th IMF component, is the lower bound of the 95% confidence interval of the j-th IMF forecast result, is the upper limit of the 95% confidence interval of the j-th IMF forecast result;

[0124] 42) Dynamic weight allocation, weight is ε=1e-6,where w j is the fusion weight of the j-th IMF component, ε is a very small constant to prevent the denominator from being 0;

[0125] 43) Weighted accumulation, the final prediction result is in, The final fused power forecast result is m, and m is the total number of IMF components. By quantifying forecast uncertainty, dynamic optimization of fusion weights is achieved, improving the reliability of results in complex scenarios. Bootstrap sampling (100 times) is performed on the forecast results of each IMF component, and a 95% confidence interval is calculated. The width of the interval reflects forecast uncertainty (such as load fluctuations caused by weather forecast deviations). The dynamic weight allocation strategy downgrades high-uncertainty components (large interval widths), suppresses abnormal interference (such as load anomalies caused by typhoons), and assigns higher weights to stable components.

[0126] A closed-loop system is formed through the above steps: the multi-level decomposition of step 2) provides component inputs adapted to different prediction models in step 3); the weight fusion of step 4) depends on the routing decision result of step 3.

[0127] The final prediction results are as follows Figure 5 As shown, the blue curve is the true value, the red curve is the prediction, the mean absolute error (MAE) is 82kW, and the root mean square error (RMSE) is 127kW.

[0128] In operation, the present invention achieves dynamic path optimization through "multi-stage adaptive noise filtering and modal decomposition" in step 2): after initial decomposition using CEEMDAN, the wavelet basis and threshold rules are dynamically selected based on the local signal-to-noise ratio, and the VMD parameters are optimized using a particle swarm algorithm. The adaptive adjustment of the secondary decomposition is achieved, and the high-frequency components are denoised using a dynamic wavelet threshold, while the low-frequency components retain trend characteristics.

[0129] Through steps 3) and 4), a dynamic fusion system driven by reinforcement learning is constructed: Step 3) An actor-critic network is designed to dynamically select the BiTCN / BiLSTM prediction branch based on characteristics such as the local variance and dominant frequency energy of the IMF component. Step 4) Bootstrap confidence intervals are introduced to calculate uncertainty weights, enabling adaptive suppression of outliers. A smoothing term is added to the reward function to avoid policy oscillation and ensure routing decision stability. A multi-objective loss function is constructed to simultaneously optimize decomposition quality and prediction accuracy.

[0130] The power load forecasting system based on reinforcement learning and dynamic mode decomposition includes:

[0131] Data acquisition and preprocessing module, used to acquire regional power load data and then preprocess the data;

[0132] The multi-stage adaptive noise filtering and modal decomposition module is used to perform complementary set empirical mode decomposition on the pre-processed load data to obtain the initial IMF component set; dynamic wavelet thresholding is then used for noise reduction, and after reconstruction, VMD is used for secondary decomposition to obtain the IMF components;

[0133] The dynamic prediction module is used to predict data of different time scales based on the IMF component by designing a dynamic routing prediction network based on deep reinforcement learning;

[0134] The fusion prediction module is used to fuse the uncertainty perception weighting with the prediction result to obtain the final prediction result.

[0135] The present invention optimizes the decomposition-forecasting collaborative process through reinforcement learning strategy to achieve accurate load forecasting.

[0136] The methods, steps, and data described herein are merely specific embodiments of the present invention and serve as a general description and illustration of the spirit of the present invention. Those skilled in the art will appreciate the numerous possibilities for variations or alternative embodiments and, guided by the spirit and principles of the present invention, may make various modifications, supplements, improvements, or substitutions. It is understood that such modifications, supplements, improvements, or substitutions are considered to be included in the present invention and do not depart from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A power load forecasting method based on reinforcement learning and dynamic mode decomposition, characterized in that: The following steps are involved: 1) Data acquisition and preprocessing: Acquire regional power load data and then preprocess the data; 2) Multi-stage adaptive noise filtering and modal decomposition: The pre-processed load data is subjected to complementary set empirical mode decomposition to obtain an initial set of IMF components. Dynamic wavelet thresholding is then used for noise reduction. After reconstruction, VMD is used for secondary decomposition to obtain the IMF components. 3) Based on the IMF component, a dynamic routing prediction network based on deep reinforcement learning is designed to predict data at different time scales; 4) The uncertainty perception weight is integrated with the prediction result to obtain the final prediction result.

2. The power load forecasting method based on reinforcement learning and dynamic mode decomposition according to claim 1 is characterized in that: In step 1), the power load data of a certain area is obtained to form an original load data set, and the data set is repaired.

3. The power load forecasting method based on reinforcement learning and dynamic mode decomposition according to claim 1 is characterized in that: Step 2) includes: 21) First, preselect the wavelet set and determine the maximum decomposition level L = log2(N)-3 according to the signal length, where N is the number of signal points; 22) Dynamic calculation of local signal-to-noise ratio, the high-frequency components after complementary set empirical mode decomposition Divide into windows of length W, with adjacent windows overlapping by 50%. For the kth window S k (t) Calculate the signal-to-noise ratio (SNR) k : Where, is the signal after preliminary denoising, S k (t) is the original signal; 23) Dynamic wavelet basis and threshold rule selection, Wavelet basis selection strategy: If SNR k <5dB, select Symlet wavelet; if 5dB≤SNR k <15dB, select Daubechics wavelet; if SNR k ≥15dB, select Coiflet wavelet; Threshold rule selection: If SNR k <5dB, using SURE threshold: Where λ is the adaptive threshold, σ is the noise standard deviation, and df(λ) is the effective degree of freedom; If SNR k ≥15dB, using Minimax threshold: 24) Layered threshold processing, for each layer of wavelet coefficient d j (j=1,2,...,L) performs dynamic thresholding: select the threshold rule according to the SNR of the current window and calculate the inter-layer adaptive threshold λ j , and then soft thresholding: Where, d j (t) is the j-th layer wavelet decomposition coefficient, λ j is the dynamic threshold of the jth layer, Coefficients after thresholding; Then perform local adaptive enhancement to enhance the amplitude of the retained wavelet coefficients. j (t)|>2λ j The coefficient of: Where α is the penalty factor; 25) Dynamic reconstruction and post-processing: using dynamically selected wavelet basis to threshold the coefficients Reconstruct and obtain the denoised signal The particle swarm algorithm is used to search for the optimal mode number K and penalty factor δ for the denoised signal. The objective function is: Perform VMD decomposition on the denoised components to obtain the final IMF set.

4. The power load forecasting method based on reinforcement learning and dynamic mode decomposition according to claim 1 is characterized in that: Step 3) includes: reinforcement learning driven dynamic routing prediction network construction and training, as follows: 31) Time series feature extraction and state definition, for each IMF component IMF j , extract local variance Main frequency energy ratio E j , autocorrelation decay rate υ j , the average error MAE in the historical window history , construct the state vector: By emphasizing that the state vector contains dynamic temporal characteristics and historical errors, the context-awareness of the strategy is enhanced; 32) Reinforcement Learning Strategy Network Design, Actor Network: Input State s j , output action probability distribution P(a j |s j ), the action space is {a BITCN , a BILSTM }; Critic network: input state s j , output value estimate V(s j ); 33) Dynamic routing decision and model reasoning, based on the action probability output by the Actor network, an ε-greedy strategy is used to select branches: Where ε is the exploration probability; Execute action a j , if a j =a BITCN , the IMF j Enter a bidirectional TCN branch; if a j =a BILSTM , the IMF j Input bidirectional LSTM branch; 34) Reward calculation and experience replay, introducing smoothing terms, and defining reward r j =-MAE j +τ·smooth term, where MAE j is the forecast error of the current IMF component, τ is the weight coefficient; the storage experience tuple (s j ,a j ,r j ,s′ j ); 35) Online optimization of the policy network, sampling batch data from the replay buffer and calculating the advantage function: A(s j )=r j +γV(s j ′)-V(s j ) (9) Where r j is the immediate reward, γ is the discount factor, V(s j ) is the value estimate of the current state, V(s j ′) is the value estimate of the next state; Update the Actor Network: Where θ is the Actor network parameter, π θ For strategic networks; Update the Critic network: 50 critic =E[(r j +γV(s j ′)-V(s j )) 2 ] (11) Where, L critic is the mean square error loss of the Critic network.

5. The power load forecasting method based on reinforcement learning and dynamic mode decomposition according to claim 1 is characterized in that: Step 4) includes: 41) Confidence interval calculation, for each IMF j The prediction results Booststrap was used to sample 100 times and the 95% confidence interval was calculated. in, is the forecast result of the j-th IMF component, is the lower bound of the 95% confidence interval of the j-th IMF forecast result, is the upper limit of the 95% confidence interval of the j-th IMF forecast result; 42) Dynamic weight allocation, weight is Among them, w j is the fusion weight of the j-th IMF component, and ε is a minimum constant; 43) Weighted accumulation, the final prediction result is in, is the final fused power forecast result, and m is the total number of IMF components.

6. The power load forecasting system based on reinforcement learning and dynamic mode decomposition is characterized by: include: Data acquisition and preprocessing module, used to acquire regional power load data and then preprocess the data; The multi-stage adaptive noise filtering and modal decomposition module is used to perform complementary set empirical mode decomposition on the pre-processed load data to obtain the initial IMF component set; dynamic wavelet thresholding is then used for noise reduction, and after reconstruction, VMD is used for secondary decomposition to obtain the IMF components; The dynamic prediction module is used to predict data of different time scales based on the IMF component by designing a dynamic routing prediction network based on deep reinforcement learning; The fusion prediction module is used to fuse the uncertainty perception weighting with the prediction result to obtain the final prediction result.

Images