Appearance
demo_vapor_pressure.py
复制以下内容并保存为 demo_vapor_pressure.py,使用方法见配套代码说明。
python
"""案例三:水的饱和蒸气压——比较训练区间内的拟合与区间外的预测。
用 0–80 ℃ 的 9 个点训练,预测 90 ℃ 和 100 ℃。三个模型对比:
A. 线性模型 P ~ T (化学里最常见的一阶近似)
B. 多层网络 P ~ T (反向传播学出来的弯曲)
C. 线性模型 ln P ~ 1/T(K) (Clausius-Clapeyron 形式)
饱和蒸气压数据(kPa)由 IAPWS 推荐的 Wagner–Pruss 饱和蒸气压方程计算,
保留 6 位有效数字。
模型 C 假设气相近似理想、液体摩尔体积可忽略,且汽化焓在该温区近似恒定。
"""
import math
from nn import MLP
# ---- 水的饱和蒸气压(IAPWS,Wagner & Pruss 2002),温度为 ITS-90 摄氏温度 ----
T_All = [0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100]
P_All = [0.611213, 1.22818, 2.33921, 4.24693, 7.38512, 12.3523,
19.9474, 31.2022, 47.4158, 70.1827, 101.418]
N_TRAIN = 9 # 前 9 个点(0–80 ℃)用于训练
T_tr, P_tr = T_All[:N_TRAIN], P_All[:N_TRAIN]
T_te, P_te = T_All[N_TRAIN:], P_All[N_TRAIN:] # 90、100 ℃ 用于外推
def mean(v):
return sum(v) / len(v)
def std(v):
m = mean(v)
return math.sqrt(sum((x - m) ** 2 for x in v) / len(v))
def rmse(a, b):
return math.sqrt(sum((x - y) ** 2 for x, y in zip(a, b)) / len(a))
def r2(y_true, y_pred):
m = mean(y_true)
ss_res = sum((x - y) ** 2 for x, y in zip(y_true, y_pred))
ss_tot = sum((x - m) ** 2 for x in y_true)
return 1.0 - ss_res / ss_tot
# 标准化统计量只用训练集计算,以免测试信息泄漏到训练中
T_mu, T_sd = mean(T_tr), std(T_tr)
P_mu, P_sd = mean(P_tr), std(P_tr)
def std_T(t):
return (t - T_mu) / T_sd
def std_P(p):
return (p - P_mu) / P_sd
def unstd_P(z):
return z * P_sd + P_mu
print("=" * 78)
print("水的饱和蒸气压:训练 0–80 ℃(9 点),外推 90 ℃ 与 100 ℃")
print("=" * 78)
# ---- A. 线性模型:P ~ T ----
Xa = [[std_T(t)] for t in T_tr]
Ya = [[std_P(p)] for p in P_tr]
A = MLP([1, 1], output_activation="linear", loss="mse", seed=0)
histA = A.train(Xa, Ya, epochs=20000, lr=0.1, batch_size=len(Xa), seed=0)
# ---- B. 多层网络:P ~ T,两个隐藏层 ----
B = MLP([1, 16, 16, 1], hidden_activation="tanh",
output_activation="linear", loss="mse", seed=0)
histB = B.train(Xa, Ya, epochs=20000, lr=0.1, batch_size=len(Xa), seed=0)
# ---- C. 物理特征:ln P ~ 1/T(K),仍然只用线性模型 ----
inv_tr = [1.0 / (t + 273.15) for t in T_tr]
inv_mu, inv_sd = mean(inv_tr), std(inv_tr)
Xc = [[(v - inv_mu) / inv_sd] for v in inv_tr]
P_REF_KPA = 1.0 # 对数参数无量纲:ln(P / P_ref),P_ref = 1 kPa
Yc = [[math.log(p / P_REF_KPA)] for p in P_tr]
C = MLP([1, 1], output_activation="linear", loss="mse", seed=0)
C.train(Xc, Yc, epochs=20000, lr=0.05, batch_size=len(Xc), seed=0)
def z_inv(t):
return (1.0 / (t + 273.15) - inv_mu) / inv_sd
print("\n训练集内(9 点,0–80 ℃)")
print(f" {'A. 线性 P ~ T':<26} R² = {r2(P_tr, [unstd_P(A.predict(x)[0]) for x in Xa]):+.4f}")
print(f" {'B. 多层网络 P ~ T':<26} R² = {r2(P_tr, [unstd_P(B.predict(x)[0]) for x in Xa]):+.4f}")
print(f" {'C. 线性 lnP ~ 1/T':<26} R² = "
f"{r2(P_tr, [P_REF_KPA * math.exp(C.predict([z_inv(t)])[0]) for t in T_tr]):+.4f}")
print("\n外推(参考值:90 ℃ 为 70.183 kPa,100 ℃ 为 101.418 kPa)")
print(f" {'T/℃':>5} {'真值':>10} {'A 线性':>10} {'B 多层网络':>12} {'C lnP~1/T':>12}")
predA, predB, predC = [], [], []
for t, p in zip(T_te, P_te):
a = unstd_P(A.predict([std_T(t)])[0])
b = unstd_P(B.predict([std_T(t)])[0])
c = P_REF_KPA * math.exp(C.predict([z_inv(t)])[0])
predA.append(a); predB.append(b); predC.append(c)
print(f" {t:>5} {p:>10.3f} {a:>10.3f} {b:>12.3f} {c:>12.3f}")
print("\n外推误差(RMSE,kPa)")
print(f" A 线性 P ~ T : {rmse(P_te, predA):7.3f}")
print(f" B 多层网络 P ~ T : {rmse(P_te, predB):7.3f}")
print(f" C 线性 lnP ~ 1/T : {rmse(P_te, predC):7.3f}")
print("\n" + "-" * 78)
print(f"模型 A 训练损失 {histA[0]:.4f} -> {histA[-1]:.6f} (2 个参数)")
print(f"模型 B 训练损失 {histB[0]:.4f} -> {histB[-1]:.6f} (321 个参数)")
print("-" * 78)