Application Study of a water conducting fracture zone height prediction model based on bp neural network optimized by genetic algorithme—Shendong mining area

Deep beneath the arid, semi-arid landscapes of northwest China lies a treasure trove of coal, a resource that has fueled the region's industrial growth for decades. However, extracting this coal comes with a steep price for the local environment. When miners dig deep into the earth to remove thick layers of coal, the rock layers above the mine do not simply stay put; they shift, crack, and collapse. This movement creates a vertical column of broken rock and open fissures that stretches all the way from the coal seam up toward the surface. Scientists call this the water-conducting fracture zone. Because these cracks act like open pipes, they allow groundwater to drain away from the surface soil and into the mine, causing the water table to drop. In a region where water is already scarce, this loss can devastate the fragile ecosystem, turning once-viable land into dust. The critical question for engineers and ecologists alike is simple yet vital: how high will these cracks reach? Knowing the exact height of this fractured zone is the only way to predict whether a mine will drain a local aquifer or leave the surface water intact.

In the Shendong mining area, known for its massive reserves and shallow coal seams, miners have long relied on old rules of thumb to guess this height. These traditional methods, often based on simple formulas derived from decades of experience, assume that the height of the damage depends mainly on how thick the coal is. But as mining technology has advanced, with machines becoming faster and working faces growing wider, these old formulas have begun to fail. They cannot account for the complex interplay of modern mining speeds, the specific angles of the coal seams, or the varying hardness of the rock layers above. To solve this, a team of researchers from the China University of Mining and Technology turned to a different kind of tool: a computer model that learns from data rather than following a fixed equation. They sought to build a system that could look at the specific conditions of a mine—how deep it is, how fast it is being dug, and how hard the surrounding rock is—and accurately predict the height of the damage zone before the mining even begins.

The researchers gathered real-world data from twenty different working faces across the Shendong area and neighboring mines. They recorded every relevant detail: the length of the mining face, the height of the coal being removed, the depth of the mine, the angle of the seam, the speed at which the machines advanced, and the proportion of hard rock in the layers above. They also measured the actual height of the water-conducting fracture zone in each location, creating a set of twenty real-life examples. To make sense of this data, they first used a statistical method to see which factors mattered most. They found that while all the factors were connected to the height of the damage, three stood out as the most influential: the length of the mining face, the speed of mining, and the height of the coal being extracted. The hardness of the rock and the depth of the mine played a role, but they were less dominant than the operational factors.

With these key factors identified, the team built a digital prediction model using a type of artificial intelligence known as a neural network. Think of a neural network as a computer system that mimics the way a human brain learns by recognizing patterns in data. However, standard neural networks can sometimes get stuck in a rut, settling for a "good enough" answer rather than the best possible one. To fix this, the researchers used a genetic algorithm, a technique inspired by natural evolution, to train the network. This process works like a digital survival of the fittest, where thousands of potential solutions are tested, the best ones are kept, and they are mixed and tweaked over many generations until the system finds the most accurate way to predict the fracture height. The result was a hybrid model, a GA-BP network, that combined the pattern-recognition power of the neural network with the optimization power of the genetic algorithm.

When the team tested their new model, the difference was stark. They compared the predictions of their new system against the actual measured heights from the mines, as well as against the old empirical formulas and a standard, unoptimized neural network. The traditional formulas were off by nearly twelve percent, and the standard neural network was even worse, missing the mark by twenty-five percent. In contrast, the new GA-BP model was remarkably precise, with an error rate of less than eight percent. To prove this wasn't just a lucky guess with the training data, the researchers took the model to a specific mine in Shangwan, where they drilled two deep boreholes to measure the actual height of the fracture zone. The model predicted a height of roughly sixty meters, while the physical measurements in the ground showed the fracture zone reached about fifty-six meters. This small gap confirmed that the model could handle real-world complexity with high reliability.

The study concludes that the height of the water-conducting fracture zone in this region is not a fixed number determined solely by the thickness of the coal, but a dynamic result of how the mine is operated. By understanding that the speed of mining and the size of the working face are the primary drivers of damage, engineers can adjust their operations to keep the fracture zone below the water table. The new model provides a practical, accurate tool for making these decisions, offering a way to balance the need for coal with the urgent need to protect the region's water resources. It represents a shift from guessing based on past experience to predicting based on a deep understanding of the specific conditions at hand, ensuring that the next generation of mining can proceed without draining the land dry.