Electrical load risk early warning method based on ARIMA model residual analysis

Abstract

The invention discloses an electrical load risk early warning method based on residual analysis of an ARIMA model, and aims to realize efficient and accurate load risk early warning through residual analysis of electrical load data. The method comprises the following steps: firstly, constructing an ARIMA model by using historical electrical load data, and predicting a load change trend; and then, detecting an unexpected or uncontrollable abnormal fluctuation condition by combining statistical analysis of a dynamic disturbance value on a model residual error. And based on residual feature extraction and threshold setting, carrying out graded early warning on potential electrical load risks. According to the method, abnormal load change can be recognized in time, and the safety and stability of power grid dispatching and operation are improved. Compared with the prior art, the method can effectively make up for the problem of insufficient load anomaly detection accuracy in a traditional method, has the characteristics of high calculation efficiency and strong applicability, and is suitable for the intelligent risk management requirements of power grid companies and large electricity users.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and in particular to an electric load risk early warning method based on time series analysis and residual detection of an ARIMA model. Background Art

[0002] With the increasing complexity of power systems and the diversification of load demands, abnormal fluctuations in electricity load pose a potential threat to grid operational safety. Traditional load forecasting methods primarily rely on statistical models (such as fixed threshold methods and autoregressive moving average models) and machine learning techniques (such as support vector regression (SVR) and long short-term memory (LSTM) networks). While these methods perform well in predicting overall trends, their real-time detection and early warning capabilities for abnormal load fluctuations are limited. For example, conventional load monitoring methods typically rely on fixed thresholds or simple trend analysis, which struggles to accurately capture the nonlinear characteristics of load fluctuations and can easily lead to false positives or omissions. Furthermore, machine learning methods are highly data-intensive and computationally complex, making them difficult to meet real-time requirements. Furthermore, existing technologies (such as CN116163807A) employ the ARIMA model for dynamic early warning of abnormal data in tunnel health monitoring. Through dynamic modeling and hierarchical threshold setting, this approach improves the accuracy of multi-step forecasts and provides dynamic early warning of abnormal data. However, this approach primarily targets tunnel structural health monitoring and lacks specificity for addressing abnormal grid load fluctuations. Furthermore, its hierarchical threshold setting, based on multiples of the standard deviation, fails to fully account for the seasonal and cyclical characteristics of grid load. Therefore, the existing technology has obvious deficiencies in the accuracy and real-time performance of load anomaly detection. There is an urgent need for an efficient and reliable load risk warning method to consider the impact of unexpected load fluctuations on the warning results, so as to reduce misjudgments and improve the adaptability of the warning system to complex load scenarios. Summary of the Invention

[0003] Aiming at the shortcomings of the existing technology in power load anomaly detection, the present invention proposes a power load risk warning method based on residual analysis of the ARIMA model (Autoregressive Integrated Moving Average Model, time series analysis model) to solve the problems of low detection accuracy and poor real-time performance of traditional methods.

[0004] The present invention adopts the following technical solutions.

[0005] The technical solution of the present invention comprises the following steps: Collect and preprocess historical electricity load data to remove outliers and noise; Build an ARIMA model based on historical data and use the model to make short-term forecasts of future loads; The residuals between the predicted values ​​and the actual observed values ​​are analyzed to extract the statistical characteristics of the residuals. The predicted value sequence is processed based on the load characteristics of the target area, fully considering unexpected load fluctuations in the target area due to uncertain demand or seasonal cycles. The residual values ​​are further calculated based on the processed predicted sequence. Each residual value in the residual value sequence is within an interval with a maximum and minimum value. Set the residual anomaly threshold and dynamically adjust the threshold based on standard deviation or probability distribution method; Triggering graded risk warnings (e.g., low, medium, and high) based on the degree and frequency of residuals exceeding thresholds; Combined with the early warning results, visual reports or real-time alarm information are generated so that the power grid dispatching center can take timely response measures.

[0006] The significant technical effects of the present invention include the following aspects: The present invention targets the seasonal and cyclical characteristics of power grid loads, or unexpected loads caused by other uncertain demands, and can further enhance the versatility and fault tolerance of abnormal situation detection, so as to adapt to temporary power load changes caused by human or sudden situations, abnormal fluctuations caused by seasons or severe weather, and improve the adaptability of the early warning system to complex scenarios.

[0007] This paper proposes a method based on dynamic standardized residual detection. After calculating the residual mean 𝜇 and standard deviation 𝜎, it introduces exponential smoothing adjustment (ESA) to perform weighted smoothing on the residuals, reduce the impact of sudden noise, and improve the stability of anomaly detection.

[0008] The present invention adopts a risk threshold setting method based on Bayesian dynamic updating, and uses the data of the past N time steps to calculate the risk threshold, so that it can be adaptively adjusted. That is, on the premise that the aforementioned residual value takes the dynamic value into account, the setting of the dynamic threshold is further improved, which solves the problem of high false alarm rate and missed alarm rate of traditional methods.

[0009] This method combines the time series prediction capabilities of the ARIMA model with the anomaly detection mechanism of residual analysis to provide real-time, accurate early warning of power load risks. Residual analysis can identify even minor load anomalies, significantly reducing false positives and false negatives. The model has low computational complexity and can quickly respond to load changes. It is applicable to power systems of all sizes and complex load scenarios.

[0010] The application of the present invention will significantly improve the safety and operating efficiency of the power system and provide strong technical support for grid operation risk management. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flowchart of the steps of the early warning method of this application; Figure 2 It is a flowchart of risk assessment and risk grading based on residuals and dynamic thresholds. DETAILED DESCRIPTION

[0012] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0013] Step 1: Data collection and preprocessing Obtain the power load data of the target area through smart meters or power monitoring systems, and perform the following processing on the data: Data cleaning: remove outliers and missing values, and fill data gaps using interpolation or smoothing; Trend and seasonality separation: Eliminate trend and cyclical features in data through differencing, logarithmic transformation, etc. to ensure data stability; Data normalization: Scale the data to the [0,1] range to facilitate subsequent model processing.

[0014] Step 2: Construction of ARIMA model Parameter identification: Use the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots to determine the autoregressive order p, difference order d, and moving average order q of the ARIMA model; Model training: Use historical data to fit the ARIMA model and optimize model parameters to ensure that it can accurately reflect load change trends; Forecast output: Use the trained ARIMA model to forecast the load in the future time period and generate the first forecast value sequence; Sequence processing: Subtract a specific constant δ from the first predicted value in the first predicted value sequence to obtain the second predicted value sequence, so that the mean of the second predicted value in the second predicted value sequence is 0. This step is a key step in locating and observing unexpected load fluctuations. It aims to generate a curve with the X-axis as the center line and the Y-axis representing the load value fluctuations by returning the middle values ​​of all predicted load values ​​to zero. This curve reflects the range of load values ​​in most scenarios to the greatest extent possible. In other words, the second predicted value fluctuates around 0 in the majority of cases. This generates the second predicted value data in a more intuitive way that facilitates subsequent model calculations. Unexpected load considerations: The data with the second predicted value range of [-a, a] in the second predicted value sequence is defined as the dynamic disturbance zone data with an output close to 0, and the data segment with continuous second predicted value data in the second predicted value sequence located in the interval of [-a, a] is defined as the dynamic disturbance segment; the load scenarios of the time series corresponding to the dynamic disturbance zone data and the dynamic disturbance segment are analyzed to obtain the dynamic disturbance zone value [-b, b] with time series characteristics; it should be noted that the absolute value of b should be a value between [0, a], and its value should be set according to the specific load characteristics of the target area. The setting of the fixed value can be determined by the specific electricity The grid management department sets this value based on load characteristics and the requirements for early warning or protection actions. When the second predicted value fluctuates slightly around the X-axis (Y = 0), the overall proportion of time periods when the second predicted value fluctuates around 0 is the highest. This allows for better capture of difficult-to-observe influences and changes in load values, such as temporary load changes caused by human intervention or emergencies, and abnormal fluctuations due to seasonality or inclement weather. In other words, these unexpected fluctuations are more likely to occur during these time periods. Locating these fluctuations within the smoothest curve can best capture unexpected load fluctuations. Obtaining the value of these dynamic disturbances through intuitive or algorithmic methods can provide a more reliable reference for risk assessment in subsequent residual analysis. Step 3: Residual Analysis Residual calculation: Calculate the difference between the predicted value and the actual observed value to obtain the residual sequence; Residual = Actual value - Predicted value Calculating the difference between a first predicted value in the first predicted value sequence and a measured value corresponding to the first predicted value to obtain a first residual sequence; summing the first residual value in the first residual sequence and the dynamic disturbance zone value [-b, b] to obtain a second residual sequence having a maximum value and a minimum value of the second residual value; Residual feature extraction: Analyze the mean μ, variance σ², standard deviation σ and distribution characteristics of the second residual sequence to determine whether it conforms to the white noise characteristics; Abnormal threshold setting: Combine the statistical characteristics of the second residual to set a dynamic threshold, for example: Threshold = μ + kσ Where μ is the mean of the second residual, σ is the standard deviation of the second residual, and k is the adjustment coefficient.

[0015] Step 4: Anomaly detection and warning Abnormal detection: When the second residual value exceeds the set threshold range, it is considered as a load abnormality; Risk grading: Risk grading is performed based on the magnitude and frequency of the second residual exceeding the threshold. The specific levels are defined as follows: Low risk: The second residual value exceeds the threshold but does not exceed 1.5σ (that is, the second residual is in the interval [μ+σ,μ+1.5σ] or [μ-1.5σ,μ-σ]), and the number of anomalies does not exceed 2 in the past 30 time steps; Medium risk: The second residual value exceeds 1.5σ but does not exceed 3σ (i.e., the second residual is in the interval [μ+1.5σ, μ+3σ] or [μ-3σ, μ-1.5σ]), or the number of anomalies in the past 30 time steps is between 3 and 6; High risk: The second residual value exceeds 3σ (i.e., the second residual is outside the interval [μ+3σ] or [μ-3σ]), or the number of anomalies exceeds 6 in the past 30 time steps.

[0016] It should be noted that since the second residual value of the present invention is an interval value having a maximum value and a minimum value, the definition of the threshold limit here should be: the interval value as a whole falls within the risk classification interval, for example, in the low risk classification, the minimum value of the second residual value is greater than μ+σ and the maximum value is less than μ+1.5σ, or the minimum value is greater than μ-1.5σ and the maximum value is less than μ-σ; the specific area definitions of medium risk and high risk are the same as above and will not be repeated here; it can be foreseen that when judging the residual value, compared with the general method, the judgment interval between the residual value and the threshold of the dynamic disturbance zone value with time series characteristics is considered, and compared with the judgment interval between the residual value and the threshold of the dynamic disturbance zone value without time series characteristics, the interval range is wider, that is, unexpected load fluctuations are taken into account, thereby achieving the effect of reducing false alarms; Real-time warning: Warning signals are sent to users through alarm systems or visualization platforms (such as dashboards), and specific abnormality descriptions are provided.

[0017] Step 5: Apply the warning results Grid dispatch optimization: Based on early warning results, dynamically adjust power generation plans and load distribution strategies to avoid supply and demand imbalances; Equipment protection measures: In response to high-risk warnings, backup equipment or distributed energy systems are activated in advance to reduce system load pressure; Data recording and review: Save warning data for subsequent model optimization and event review analysis.

[0018] To verify the effectiveness of the method of the present invention, this section conducts case analysis through actual power load data and conducts comparative experiments with traditional methods to demonstrate the advantages of the present invention in anomaly detection and risk warning.

[0019] The experimental data is the power grid load data of a certain region from October 2023 to November 2024, with a data sampling frequency of 15 minutes, totaling 35,040 data points. The data preprocessing steps include: (1) Z-score standardization is used to eliminate the dimension effect; (2) Use the IQR (interquartile range) method to eliminate extreme outliers; (3) The first-order difference method is used to eliminate the trend and ensure the stationarity of the time series.

[0020] Risk warning calculation process based on the method of the present invention ARIMA model construction and short-term forecasting: (1) The ADF (Augmented Dickey-Fuller) test was used to determine the stationarity of the data and to determine the optimal difference order d = 1.

[0021] (2) PACF (partial autocorrelation function) and ACF (autocorrelation function) were used to determine the ARIMA model order (p, d, q) combination, and ARIMA (2, 1, 2) was finally selected as the best model.

[0022] (3) Use data from October 2023 to September 2024 for training and data from October to November 2024 for testing, and use the trained ARIMA model for short-term load forecasting.

[0023] Residual analysis and dynamic threshold setting: (1) Calculate the residual sequence between the predicted value and the actual value, and calculate its mean μ and standard deviation σ.

[0024] (2) The weighted residual is calculated using the exponential smoothing adjustment (ESA) method, with the weight set to α = 0.8. The smoothing formula is as follows: R t ′=αR t +(1-α)R t-1 ′where R t is the residual at the current moment, R t ′ is the residual after smoothing.

[0025] (3) Dynamic threshold setting: Based on the Bayesian dynamic update method, the risk threshold is recalculated every 100 time steps to make it adaptively adjusted.

[0026] Risk classification warning Based on the magnitude and frequency of residuals exceeding the threshold, the risk grading standards are set as follows: Low risk (0.674σ ≤ residual <1.28σ, and the proportion of outliers ≤5%) Medium risk (1.28σ ≤ residual <1.96σ, and the proportion of outliers >5% and ≤15%) High risk (residual ≥ 1.96σ, or outlier percentage > 15%) When the residuals in a certain time period meet the risk conditions, the system generates an early warning report and displays the load anomaly in a visual manner.

[0027] Comparative analysis of experimental results In order to evaluate the effectiveness of the method of the present invention, comparative experiments were carried out with the fixed threshold method (traditional empirical method), the load anomaly detection method based on K-means clustering (comparison method 1), and the load forecast anomaly detection method based on LSTM (comparison method 2).

[0028] (1) Comparison of anomaly detection accuracy Table 1

[0029]

[0030] (2) False Positive Rate (FPR) comparison Table 2

[0031]

[0032] (3) Comparison of calculation time (unit: seconds) Table 3

[0033]

[0034] Experimental Conclusion The proposed method improves the anomaly detection accuracy by 4-10 percentage points, with an F1-score of 90.7%, which is superior to existing methods. The false alarm rate is reduced to 5.2%, which is more than 60% lower than traditional methods, avoiding unnecessary warning interference; The computational efficiency is better. Compared with the LSTM method, the training time is reduced by 83.7% and the prediction time is reduced by 68.4%. It is suitable for real-time load warning.

[0035] Compared with the fixed threshold method: the traditional method has a fixed threshold and cannot adapt to the dynamic changes of the grid load, while the dynamic threshold adjustment mechanism of the present invention can change with the load characteristics and reduce false alarms.

[0036] Compared with the K-means clustering method: the K-means method requires manual setting of the number of categories and is sensitive to data distribution, while the present invention adaptively adjusts the threshold based on residual statistical characteristics and is more robust.

[0037] Compared with the LSTM prediction method: Although LSTM has the ability to predict time series, it has high computational complexity. The present invention is based on ARIMA+ dynamic residual analysis, which reduces computational overhead while ensuring high accuracy and is suitable for large-scale power grid load warning applications.

[0038] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A power load risk warning method based on ARIMA model residual analysis, characterized in that: The following steps are involved: Obtain the electricity load data of the target area and preprocess the data; Construct an ARIMA model and determine the autoregressive order, difference order, and moving average order of the ARIMA model; Fitting an ARIMA model using historical data from a first specific time period of the target area, optimizing model parameters, and obtaining a trained ARIMA model; The trained ARIMA model is used to predict the load in the second specific time period to generate a first prediction value sequence; Subtracting a specific constant δ from the first predicted value in the first predicted value sequence to obtain a second predicted value sequence, so that the mean of the second predicted values ​​in the second predicted value sequence is 0; The data in the second prediction value sequence with the second prediction value range of [-a, a] is defined as dynamic disturbance zone data with an output close to 0, and the data segment in the second prediction value sequence with continuous second prediction value data in the interval of [-a, a] is defined as a dynamic disturbance segment. The load scenario of the time series corresponding to the dynamic disturbance zone data and the dynamic disturbance segment is analyzed to obtain the dynamic disturbance zone value [-b, b] with time series characteristics. Calculating the difference between the first predicted value in the first predicted value sequence and the measured value corresponding to the first predicted value time series to obtain a first residual sequence; summing the first residual value in the first residual sequence and the dynamic disturbance zone value [-b, b] to obtain a second residual sequence having a maximum value and a minimum value of the second residual value; Analyze the distribution characteristics of the second residual sequence and set a dynamic threshold; When the second residual value exceeds the set threshold range, it is considered as load abnormality; A risk classification warning is issued based on the amplitude and frequency of the second residual value exceeding the threshold.

2. The early warning method according to claim 1, characterized in that: The step of determining the autoregressive order, the differencing order, and the moving average order of the ARIMA model includes: using an ADF (Augmented Dickey-Fuller) test to judge the stationarity of the data and determine the optimal differencing order d; using an autocorrelation function (ACF) and a partial autocorrelation function (PACF) graph to determine the autoregressive order p and the moving average order q of the ARIMA model.

3. The early warning method according to claim 2, characterized in that: Analyzing the distribution characteristics of the second residual sequence and setting the dynamic threshold comprises: The mean μ, variance σ² and standard deviation σ of the second residual sequence are calculated; the weighted residual is calculated using the exponential smoothing adjustment (ESA) method, with the weight set to α=0.

8. The smoothing formula is as follows: R t′ =αR t +(1-α)R t-1′ Among them, Rt is the residual at the current moment, and Rt′ is the residual after smoothing. Based on the Bayesian dynamic update method, the dynamic threshold is recalculated every 100 time steps to adaptively adjust the dynamic threshold.

4. The early warning method according to claim 3, characterized in that: The analyzing the distribution characteristics of the second residual sequence and setting the dynamic threshold further includes: Dynamic threshold = μ + kσ Among them, μ is the residual mean, σ is the residual standard deviation, and k is the adjustment coefficient.

5. The early warning method according to claim 3, characterized in that: The risk classification warning according to the magnitude and frequency of the second residual exceeding the threshold value includes: The risk grading standards are set as follows: Low risk: 0.674σ ≤ second residual value < 1.28σ, and the proportion of outliers ≤ 5%; Medium risk: 1.28σ ≤ second residual value < 1.96σ, and the proportion of outliers is >5% and ≤15%; High risk: The second residual value is ≥ 1.96σ, or the proportion of outliers is > 15%.

6. The early warning method according to claim 4, characterized in that: The risk classification warning according to the magnitude and frequency of the second residual exceeding the threshold value includes: The risk grading standards are set as follows: Low risk: The residual value exceeds the threshold but does not exceed 1.5σ (i.e., the residual is in the interval [μ+σ, μ+1.5σ] or [μ-1.5σ, μ-σ]), and the number of anomalies does not exceed 2 in the past 30 time steps; Medium risk: The residual value exceeds 1.5σ but does not exceed 3σ (i.e., the residual is in the interval [μ+1.5σ, μ+3σ] or [μ-3σ, μ-1.5σ]), or the number of anomalies in the past 30 time steps is between 3 and 6; High risk: The residual value exceeds 3σ (i.e., the residual is outside the interval [μ+3σ] or [μ-3σ]), or the number of anomalies exceeds 6 in the past 30 time steps.

7. The early warning method according to claim 1, characterized in that: The data preprocessing comprises: Z-score standardization is used to eliminate the dimension effect; the IQR (interquartile range) method is used to eliminate extreme outliers; the first-order difference method is used to eliminate trends and ensure the stationarity of the time series.

8. The early warning method according to claim 1, characterized in that: For the dynamic disturbance zone data [-a, a] and the dynamic disturbance zone value [-b, b], the absolute value of b is a value between [0, a].

Images