When Power Flow Models Go Off the Rails

Links table

Abstract and 1. Introduction

2. Evaluation methods

3. Review the current experiments

4. Circular and application assessments and 4.1. Prediction and response circulation

4.2. Applications for multiple -written situations and 4.3. Zero predict the ability of the application

4.4. Continuous prediction and 4.5. Normalization

5. numerical assessments and 5.1. Experience settings

5.2. Overview of the evaluation

5.3. Failure evaluation

5.4. Accuracy

5.5. Efficiency evaluation

6. Open questions

7. Conclusion

Approach a and references

4, generalization and application assessments

4.1. Prediction and response circulation

While PPFL methods are often restricted in their choice of predictions and responses due to specific material fixtures, DPFL roads generally provide a more flexible frame. This flexibility indicates the possibility of following DPFL methods to accommodate the known known variables (including 𝑃 and 𝑄 of PQ buses, 𝑉 of stagnation and photovoltaic jets, and 𝜃 of the stagnation bus) as predictors, and unknown variables (including 𝑃 from the stagnation bus , 𝑄 from stagnation and jawwah buses, 𝑉 from PQ buses, 𝜃 from PQ and PV buses, and all active/interactive flows, IE, PF, PT, QF, QT). However, due to various factors, not all DPFL methods can achieve this level of generalization:

• For methods of PLS_BDL and PLS_BDLY2, using 𝑉, 𝑃 and 𝑄 Parents are also implemented through the package strategy [6] They adopt. This strategy, designed to address differences in bus types, restricts the nature of the choice of predictors by identifying known and pre -known variables.

• LCP_Box and LCP_JGD methods are merged with energy flows by formulating construction restrictions [6]. Within this Jacobi matrix, 𝑃 and 𝑄 are dealt with as known variables, while 𝑉 and 𝜃 are unknown variables.

• The LCP_Cou method is specifically designed to estimate the values ​​of the branch’s flows in line with the voltage and peripheral angles as predictive variables [6]. Thus, the method can only be used 𝑉 and 𝜃 as predictions, PF, PT, QF and QT responses.

• For DC_LS and DLPF_C methods, because they integrate DC and DLPF models, respectively, in the framework of their work [6]They must align the selection of predictors and responses with the basic material models they adopt.

Note that the flexibility restricted to the choice of predictors and responses leads to noticeable restrictions. First, these methods may not benefit from all the known data available for models training, which leads to the loss of potential information. For example, the DC_LS method is only used measurements 𝑃, with a large amount of known voltage data. Second, the ability to predict the unknown variables can be restricted using the developed linear model. For example, the LCP_Box method is limited to the expense of the branch flow values, which leads to a very limited functional range.

4.2. Application on multiple -written situations

The regular small squares method are struggled with multiple linear lines [6]. LS_cls and LS_RC methods share this restriction, as they are all based on the normal small squares frame. In addition, LS_SVD and LS_TOL are also affected by this problem, as it has been discussed in [6]. Subsequent experiences will show a number of their restrictions in this context.

4.3. Zero predict the ability of the application

The issue of zero predictions arises when some variable measurements known in the training data set are constantly zero. Typical examples include inserting the corner of the conveyor in the prediction group, which is usually assigned to be zero and remains zero all the time. Other cases may include PQ buses where active/interactive energy consumption is zero during the measurement period. This position leads to zero columns in the prediction data collection matrix (where the columns represent different variables, and the rows represent individual measurements). Not all DPFL methods can deal with these zero columns effectively:

• Methods that depend on regular small squares, including LS_Cls and LS_RC, face difficulties with zero predictions, as these zero columns make the gram matrix for the non -predictable predictable matrix, which leads to the failure of these methods.

4.4. Continuing prediction

The problem of fixed prediction extends beyond the zero prediction problem, which occurs when the measurements of some variables in the training data set remain constant, not necessarily zero. A typical example is the fixed peripheral voltage in PV buses, which usually remain fixed during the measurement period, which leads to columns with fixed value in the prediction matrix.

4.5. Normalization

As detailed in [6]Integration of physical knowledge into DPFL methods can become a problem with data groups that have been normalized through contrast techniques, such as energy normalization, where each variable is normalized independently. This independent normalization disrupts the connected physical relationships between the variables, such as this actress in the Jacobian Matrish or through conjugation relationships, and methods of presentation such as RR_VCS, LCP_BOX, LCP_Cou, LCP_JGD, DC_LS, and DLPF_C are not connected.